Changelog

Release notes

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Releases 1.0.0–3.0.0 below were reconstructed retrospectively (this changelog was introduced after 3.0.0 shipped), organized by capability rather than by individual commit. The capability ladder matches VERSION and UPGRADE.md.

Version 4.34.1, 8/14/2026

Secret-guard output readability (PATCH). Field feedback from the maintainer's own regression test of the v4.34.0 pre-commit guard: every finding line repeated the same advisory tail, burying the findings. Guidance now appears once per run, not once per finding.

Changed

• [secret-material] finding lines are findings only — they end at (N hit(s), first at line N); the redact/rotate/history guidance moved to a single consolidated line per consumer: the pre-commit hook's -> fix it footer (now preceded by a blank separator line and carrying the shared/rotation/history wording), a one-line trailer in --scan-files mode, and a one-line trailer after the warnings in a full memory-lint run. Both runtimes at parity; 49 mirror tests unchanged and passing; the hook's grep contract is unaffected (the guidance line carries no [secret-material] tag).

Targets: re-copy the memory-lint built-in + .githooks/pre-commit, stamp.

Version 4.34.0, 8/14/2026

Pre-commit secret guard — the layer that prevents instead of detects (MINOR). The v4.33.x arc left one timing gap: agents get the redaction rule at write time and every push meets the CI floor — but a human teammate committing directly meets no automated check until the remote already has the secret, where redaction is no longer un-leaking and rotation begins. And the field incident's origin was never a memory file: the credentials entered the repo inside a Postman JSON and an OpenShift YAML, then contaminated a session log when a dry-run rendered them. The new committed .githooks/pre-commit closes the gap at the only placement that can — before the commit exists — and covers both surfaces.

Added

• .githooks/pre-commit — scans the staged content (the index via git show :path, not the worktree: exactly what this commit would publish; only what THIS commit stages, so a pre-existing finding elsewhere never gates an unrelated commit) of:
• memory/**.md — the full [secret-material] profile (credentials + PII);
• config files — .json / .yml / .yaml / .properties / .toml / .ini / .env and .env.* anywhere in the repo — credential-class checks only (token shapes, key assignments, Authorization headers, private keys): config files legitimately carry contact emails and paths; credential material never is. Findings print the linter's non-echoing report plus redaction / waiver / rotation guidance. Enforcing by default (maintainer decision): findings block the commit — the deliberate, scoped exception to the trigger layer's advisory doctrine, because secrets carry irreversible after-the-fact cost and an advisory default lets an inattentive commit ride to the remote. Opt down to warn-only with AGENT_MEMORY_SECRET_GUARD=advisory (env or git config agent-memory.secretguard advisory); one-off bypass git commit --no-verify; merge commits skipped; python3→node fallback (with neither, it skips with a note). Rides the existing core.hooksPath activation — zero new setup on initialized clones; a never-initialized clone falls through to the CI floor.
• memory-lint --scan-files FILE... (both runtimes): the standalone credential-class scan powering the hook and the CI wrappers — paths reported as given, exit 1 on findings, values never echoed.
• Changed-config secret scan on the CI floor (all three forge wrappers): the push-time sibling of the hook, scanning the push/PR's changed config files with the same advisory / STRICT semantics — catches --no-verify bypasses and never-initialized clones.
• .agent/secret-scan-ignore (committed, human-audited; installed as a commented stub): shell-glob-per-line exemptions for formats that can't carry the inline waiver tag (JSON has no comments; a .properties same-line comment corrupts the value). Config files only — memory/ is never exempt. Honored by the hook and all three CI wrappers.
• Detector refinements from a 661-file live-corpus probe (every committed json/yml/yaml/properties file across two production repos — 6 findings, all false positives, → 0 after): single-brace {PLACEHOLDER} and GitHub-Actions ${{ … }} template forms; demo/test join the placeholder words; a template default that is itself provably a placeholder is safe (${DEMO_PEER_TOKEN:demo}) while credential-shaped fallbacks keep flagging (the v4.33.4 rule); dotted lowercase route references (v1.basic.auth) are exempt from the Authorization check; ; joins the value delimiters (JAAS terminators); and a Postman split-pair pattern ("key": "client_secret", "value": "…") covers the literal incident artifact class. Mirror suites grow 46 → 49 tests per runtime, byte parity held.

Changed

• Trigger-layer docs enumerate both hooks, both surfaces, and the refined doctrine — advisory by default; the pre-commit secret guard alone enforces: AGENTS.md root
• template, ENABLE.md (Step 5e ignore-stub install, Step 6 hooks + CI floor descriptions, both-hooks chmod line), .githooks/README.md + init.sh, SKILL.md body (--scan-files), docs/DESIGN-ritual-triggers.md (v4.34.0 amendment — its §7.3 fork had recommended an opt-in pre-commit half back at v4.19.0), the whitepaper ×2, and the docs-site reliability + built-in-skills pages.

No memory-file shape change; SKILL.md frontmatter description unchanged → adapters untouched. Verified end to end: 49/49 both runtimes; hook exercised across fourteen scratch-repo paths (default-block / advisory-opt-down via env and git-config / waiver / ignore-file / clean-config / index-vs-worktree / fast-path / node-only / --no-verify / .env / Postman-JSON / memory-regression); live corpus scans clean through the real CLI in both runtimes. Targets: copy the new hook + ignore stub, re-copy the built-in + hook docs + forge CI file, chmod +x, stamp.

Version 4.33.4, 8/13/2026

[secret-material] — reject non-empty template defaults (PATCH). Post-release review found that v4.33.3's broad ${…} placeholder exemption also trusted literal fallback values. A line such as client_secret=${CLIENT_SECRET:-RealSecret123} could therefore evade assignment detection.

Fixed

• ${NAME}, ${NAME:}, and dotted references such as ${vault.paths.kafka} remain safe placeholders. Any non-empty fallback is no longer exempt, so client_secret=${CLIENT_SECRET:-RealSecret123} flags without echoing the value.
• Python and Node remain at output parity; mirror suites pass 46 tests each.

No memory-file shape change; SKILL.md description unchanged, so adapters are untouched. Targets re-copy the tool-managed memory-lint built-in and stamp.

Version 4.33.3, 8/13/2026

[secret-material] security-review hardening (PATCH). A fresh-context security review found that the detector was useful but not a reliable enforcement control: forge wrappers never passed --strict, the v4.33.1 ALL-CAPS exemption trusted uppercase secrets globally, common quoted/Authorization forms could bypass assignment detection, and Mode C did not require migrated-history triage. This patch closes all four findings while preserving the advisory-by-default doctrine.

Security

• Forge CI enforcement fixed: GitHub, GitLab, and Azure wrappers now invoke memory-lint --strict, making warnings observable to the wrapper. The wrapper still converts findings to its native advisory result by default; AGENT_MEMORY_STRICT=1 now genuinely blocks.
• Uppercase-secret bypass closed: ALL-CAPS enum values are exempt only when the key names an enum dimension (source, type, mode, mechanism, or strategy). Uppercase passwords, client secrets, and opaque tokens on secret-bearing keys flag.
• Pasted-output coverage expanded: quoted JSON/YAML credential keys and Authorization/Proxy-Authorization headers flag without echoing their values. Placeholder recognition is full-value anchored, so embedding dummy, sample, or similar text no longer launders a credential; template/redaction forms remain safe.
• Mode C migration hardened: MIGRATE.md requires redaction during conversion, and ENABLE.md verification requires [secret-material] triage before the enable commit.
• Fold-in (pre-tag, from the post-merge live-target review): the tightened template recognition initially accepted only ${IDENTIFIER}, re-flagging standard default-value and dotted env-var forms (${REDIS_PASSWORD:}, ${VAR:-x}, ${a.b.c}) — including a target repo's own documented safe pattern for secrets (two live false positives on mercury-composable). The ${…} alternative now accepts any brace-delimited reference; recognition stays full-value anchored, so a partially-literal value still flags. The motivating line is a verbatim mirror fixture.

Both runtimes remain at output parity; mirror suites now pass 46 tests each. No memory-file shape change; SKILL.md description unchanged, so adapters are untouched. Targets re-copy the built-in and their forge CI file, merge the Mode C redaction rule if they use migration, then stamp.

Version 4.33.2, 8/13/2026

[secret-material] — backtick is a value delimiter (PATCH). The v4.33.1 enum-constant exclusion missed the form the motivating field line actually used: markdown inline code. In `key=VALUE` the closing backtick rode into the captured value (OAUTHBEARER`), so the ALL-CAPS rule didn't match and the false positive survived. Caught minutes after release by the 4.33.1 rung's own verify step against the live target. Every scanned memory surface is markdown — the assignment pattern now treats backticks like quotes.

Fixed

• memory-lint credential-assignment pattern: backtick joins the quote/value delimiter set (`key=VALUE` captures VALUE, not VALUE`) — both runtimes at parity; the mirror enum test now uses the exact field line form plus a backticked real-secret negative control (45 each). Verified against the live target: zero [secret-material] hits with no waiver.

Changed

• PR/MR description templates (this repo's own + all three forge variants): the rendered <sub> convention-explainer footer became an HTML comment — it still guides authors in the pre-filled body, but never renders in a created PR/MR. From maintainer feedback on this release's own PR: the explanation is template guidance, not PR content.

No memory-file shape change; SKILL.md description unchanged → adapters untouched. Targets: re-copy the built-in + stamp — the mercury-composable waiver becomes genuinely droppable here (4.33.1's rung said it; 4.33.2 makes it true) — and optionally re-copy the forge description template if uncustomized.

Version 4.33.1, 8/13/2026

[secret-material] — ALL-CAPS enum constants are not credentials (PATCH). Check 10's first field contact (the 2026-08-13 Mode B upgrades) produced exactly one finding across two production repos — and it was a false positive: a session log documenting Confluent's bearer.auth.credentials.source property, whose value there is the enum constant OAUTHBEARER — a source type, not a credential. The credential-assignment pattern now recognizes ALL-CAPS identifiers (OAUTHBEARER, SASL_SSL, STATIC_TOKEN, …) as config constants: real credentials carry mixed case/symbols, and the famous uppercase-only shapes (e.g. AWS access-key ids) stay covered by the value-shape patterns independently. The advisory stays signal, not noise — fix the detector, don't sprinkle waivers through client repos.

Fixed

• memory-lint _is_placeholder_value / is_placeholder_value: values matching ^[A-Z][A-Z0-9_]{2,}$ are treated as config enum constants, not secrets — both runtimes at parity, mirror test added (45 each).
• Targets that waived this exact FP class (mercury-composable, session 2026-07-09-212417.md) may drop the waiver tag after upgrading — noted in the rung.

No memory-file shape change; SKILL.md description unchanged → adapters untouched.

Version 4.33.0, 8/13/2026

Session-log secret redaction — ritual rule + memory-lint [secret-material] advisory (MINOR). A client-side DLP scanner caught a live OAuth client secret in a committed session log: the agent had pasted smoke-test output verbatim, and nothing in the protocol or tooling stood between a rendered credential and git push. Session logs and continuity are committed and shared by design — so the ritual now carries an explicit redaction rule, and memory-lint gains a deterministic backstop for the detectable shapes. Doc rule first, observational lint second, no mutation — the tool warns; the human redacts and rotates.

Added

• Redaction rule in the after-session ritual (AGENTS.md root + template; echoed in ENABLE.md Step 5c and the schema's session-file section): never write secrets or PII into memory/ — redact pasted command/smoke-test output to (REDACTED) before persisting (template placeholders like ${VAR} are already safe); a secret that was committed anyway is exposed — redaction is not un-leaking; rotate it, and git history cleanup is a separate, human-led step. Redaction is the one sanctioned edit to an otherwise-immutable session log.
• memory-lint check (10) [secret-material] (both runtimes at parity + mirror tests, 44 each): scans every committed memory surface — memory/*.md, memory/sessions/, and memory/archive/ (where pasted output lives; the conflict-marker check's session exclusion deliberately does NOT apply here) — for known token shapes (AWS / GitHub / GitLab / Slack / Google, private-key blocks, JWTs), credential-key assignments with literal values (the rendered-JAAS class; placeholders and (REDACTED) recognized as safe, number/date shapes excluded), emails (noreply / git@ / example forms excluded), SSN and Luhn-verified payment-card shapes, and absolute home paths (write ~). Advisory WARN (--strict / CI AGENT_MEMORY_STRICT=1 gate it red); the report never echoes the matched value — a lint line quoting the secret would amplify the leak into terminals and CI logs; lint:allow-secret-material on a line waives a deliberately-quoted example.
• Dogfooded on this repo (145 session logs + full archive): two real pre-existing home-path hits surfaced (mock-path quotes in old cleanup logs — waived in place with the new marker); zero credential/PII noise across all technical content.

Changed

• memory-lint SKILL.md (check list + description — description change → adapters re-synced).
• Lockstep: VERSION → 4.33.0; CHANGELOG; README (row + 10-cap trim); UPGRADE.md (row + 4.32.1 → 4.33.0 rung); docs site (reference/built-in-skills.md — ten checks). No memory-file shape change.

Version 4.32.1, 8/6/2026

Mode A last_session contradiction fix (PATCH). A real Mode A enable (2026-08-06), confirmed by an adversarial protocol audit, surfaced a tension ENABLE.md had carried since Step 5c was added: Step 5b still said a non-migrated enable leaves last_session: (none yet), while Step 5c mandates writing a first enable session log for every fresh enable — so "(none yet)" was false the moment the enable completed, and it blinded the multi-agent continuity check (AGENTS.md reads last_session to decide whether another agent family worked last). Step 5b's own footer bullet already pointed the seeded facts' origin at the 5c log; last_session now points there too.

Fixed

• ENABLE.md Step 5b: a non-migrated enable now points last_session at the Step 5c enable log — <today> | agent: <your agent name> (<the 5c log's filename stem>) — filled when that log is written (the same moment its stem becomes the seeded facts' origin). The migrated-history branch (Mode C) is unchanged. Step 5c states the back-reference.
• templates/memory/continuity.md: the literal (none yet) seed became a {{LAST_SESSION}} placeholder — the value is mode-dependent and always filled (the {{BOOTSTRAP_IMPORTS}} pattern, v4.29.1).
• templates/.agent/schema.md: the last_session line now marks (none yet) as legacy (pre-4.32.1 enables only).
• examples/rust-event-bus/: the fixture stays unedited by design — it truthfully predates Step 5c; its ENABLE_OUTPUT.md header now says so, so its (none yet) isn't read as current behavior.

No memory-file shape change (the schema touch is a value-domain note, not a field change); skills/adapters unchanged. Existing targets: re-copy the schema + stamp; optionally point a still-(none yet) last_session at the newest session log if one exists.

Version 4.32.0, 7/27/2026

Azure DevOps forge support — own-pipeline ritual floor + PR template (MINOR). The third forge, and unlike Bitbucket (investigated and deferred the same week) this one has a real field installation behind it. Azure DevOps verified as structurally different from both shipped forges: the committed artifacts map cleanly, but activation is not file-driven — a pipeline is a resource that must be bound to the YAML once, and Azure Repos ignores the YAML pr: key entirely (PR validation lives in branch policies). The design leans into what the forge does well and is honest about what the tool cannot do for you.

Added

• Own-pipeline CI floor: templates/.azuredevops/agent-memory-ci.yml — a complete, self-contained pipeline, so an existing azure-pipelines.yml is never touched (no include splicing, no stage hazards; one repo can carry many pipelines). Same two checks as the other forges; the best advisory semantics of the set — ##vso[task.logissue type=warning] annotations
• task.complete result=SucceededWithIssues give a native "partially succeeded" (orange) tri-state; AGENT_MEMORY_STRICT=1 fails the run; fetchDepth: 0 (new ADO pipelines default to shallow=1); PR-validation builds diff against HEAD^1 (the merge commit's target parent).
• PR template: templates/.azuredevops/pull_request_template.md — auto-applies to new PRs (default-branch-read; .azuredevops/ is first in ADO's search order; 4000-char description cap noted in the footer).
• Forge detection extended (ENABLE.md): dev.azure.com / *.visualstudio.com remotes + azure-pipelines.yml / .azuredevops/ signals; Step 6/8/9 + write-scope gained the third branch. Unknown forge now installs the GitHub + GitLab sets only — the ADO set requires positive detection, since its pipeline needs activation and a blind install is pure noise.
• The honest limit, stated everywhere it matters: enable reports the one-time az pipelines create … --skip-first-run command (Contributors-level) and runs it only at the user's explicit direction; PR-time validation is documented as the optional Build Validation branch policy (admin; "Optional" mode = notify-only); Microsoft-hosted agents need parallelism — the free grant (Microsoft's request form) or paid parallel jobs via a linked Azure subscription — or a self-hosted agent.

Changed

• Squash-merge guidance gains the third branch (AGENTS.md root + template): Azure DevOps drops trailers like GitLab, but has no message-template mechanism at all — the durable record is the PR-description footer; re-add the trailer via "Customize merge commit message" at completion.
• Forge-qualified claims updated across .githooks/README.md, .githooks/init.sh, AGENTS.md ×2, docs/DESIGN-ritual-triggers.md (a v4.32.0 status-header amendment), docs/optional-ritual-hook.md, and the six docs-site pages.
• Lockstep: VERSION → 4.32.0; CHANGELOG; README (row + 10-cap + file tree); UPGRADE.md (row + 4.31.0 → 4.32.0 rung). No memory-file shape change; skills/adapters unchanged. Bitbucket remains a trigger-gated backlog thread.

Version 4.31.0, 7/27/2026

GitLab forge support — forge-aware ritual floor + MR template (MINOR). A GitLab-hosted installation reported that GitLab ignores .github/ entirely (verified: no compatibility shim — templates, CODEOWNERS, and the importer all skip it). A 5-agent gap analysis found exactly two installed artifacts die there — but one is load-bearing: the CI floor never runs, so a fresh clone where no agent activated the git hook has no ritual backstop at all (the precise gap v4.19.0 existed to close), and the What/Why PR template never seeds. Root cause: "vendor-neutral" had quietly conflated AI vendor with hosting forge. Everything else under .github/ (copilot-instructions, skills adapters) is read by local tooling and was never broken — the fix deliberately leaves those untouched on every forge.

Added

• Forge detection (ENABLE.md Step 4): remote URL + .gitlab-ci.yml/.gitlab/ signals; .gitlab-ci.yml is now a recognized CI structure signal. Unknown forge → install both sets (each forge ignores the other's files — additive-safe).
• GitLab CI floor: templates/.gitlab/agent-memory-ci.yml — the same two checks as the GitHub workflow (memory-lint + advisory session-log presence), advisory via allow_failure: exit_codes: [42] (orange warning, never blocks; AGENT_MEMORY_STRICT=1 exits 1 and gates), image: python:3 (git needed in-container), GIT_DEPTH: 0 for the diff-based check. Root wiring is additive-safe: no .gitlab-ci.yml → install templates/.gitlab-ci.yml verbatim (include + the canonical workflow:rules duplicate-pipeline guard, which GitLab requires in the root file for MR pipelines); pre-existing .gitlab-ci.yml → add-only include: entry, never workflow:rules (that would change when the repo's own jobs run; the job then rides branch-push pipelines — still a real floor).
• GitLab MR template: templates/.gitlab/merge_request_templates/Default.md (auto-applies to new MRs, all tiers) — same What/Why + Co-Authored-By footer, plus the GitLab squash note.
• Step 6/8/9 + the write-scope notes gained forge-conditional branches; the enable report now names the detected forge and, for self-managed GitLab, the runner prerequisite.

Changed

• Squash-merge co-author guidance is forge-aware (AGENTS.md root + template): the failure mode inverts across forges — GitHub compounds trailers (keep one, trim repeats; the v4.28.x rule), GitLab drops them (default squash message is the MR title only) — so on GitLab the guidance is to make the trailer survive: re-add it when editing the squash message at merge, or set %{all_commits} in the project squash template (it embeds full commit messages incl. body trailers; GitHub-like compounding, so the dedup rule applies). %{co_authored_by} does not carry body trailers — it credits commit authors only (verified against GitLab's message generator; the adversarial pre-ship review caught the earlier misprescription). "Pull request" wording is now "pull/merge request" where installed.
• Every "CI is the zero-config backstop" claim is forge-qualified — .githooks/README.md, REVIEW.md, AGENTS.md (root + template), docs/DESIGN-ritual-triggers.md (a status-header amendment: the vendor-neutrality argument now names the forge dimension explicitly), docs/optional-ritual-hook.md, and six docs-site pages (getting-started, index, reliability, enable-a-repo, built-in-skills, whitepaper). Honest limit stated where it bites: GitLab.com runners are zero-config; a self-managed GitLab needs an admin-registered runner.
• Lockstep: VERSION → 4.31.0; CHANGELOG; README row; UPGRADE.md row + 4.30.0 → 4.31.0 rung. No memory-file shape change; skills/adapters unchanged. Non-goals kept lean: no CODEOWNERS, no issue templates, no further forges yet — but the forge seam is where they'd slot in.

Version 4.30.0, 7/15/2026

Stack-aware .gitignore build-output seed (MINOR). A greenfield field case (mercury, a Rust port started from an empty repo, enabled at v4.29.1): the installed .gitignore is deliberately scoped to AI infrastructure, so when the stack landed after enable, the first cargo build polluted git status until the user hand-added target/. The gap splits by when the stack is knowable: brownfield enables can seed at Step 7; greenfield enables can only carry the action forward to the moment the stack lands — and the agent working in the repo then is the runtime that applies it.

Why: the enable advisory already promises "add-only edits to .gitignore"; a repo whose first build dirties git status undercuts the faithful-enablement goal for the price of a six-row table. Deliberately bounded — an explicit non-goal note keeps it a minimal seed, never a gitignore manager.

Added

• ENABLE.md Step 7 → "Stack-aware build-output seed": when Step 4 detected a primary language/stack, append only the missing canonical build-output entries under a second, separately-scoped sentinel (# === agent-memory: build output (stack-aware seed …) ===) — add-only, de-duplicating against the whole file (a no-op for repos that already ignore them). Seed table: Rust target/; Node/JS/TS node_modules/, dist/; Python .venv/, venv/; Java/Kotlin target/, build/, *.class; .NET bin/, obj/; Go none (builds out of tree).
• ENABLE.md Step 5b: greenfield enables seed a - [ ] Greenfield — no code yet Open Thread listing what to record when the first code lands — stack, conventions, invariants, and the stack's build-output ignores (the Step 7 table, applied at that moment).
• ENABLE.md Step 8 (verify) checks the detected stack's build output is ignored (pre-existing or seeded); Step 9 report gains a greenfield next-steps line.

Notes

• Operator-side only (ENABLE.md): no template, memory-file, skill, or adapter change — templates/.gitignore stays AI-infrastructure-scoped by design.
• Already-enabled repos: nothing required; a future Mode B upgrade may apply the Step 7 seed additively (see the 4.29.1 → 4.30.0 rung).

Version 4.29.1, 7/12/2026

Template import blocks become a {{BOOTSTRAP_IMPORTS}} placeholder (PATCH). Cross-vendor dogfooding of v4.29.0 (a GitHub Copilot assessment, independently corroborated live on Claude Code) found a tool-repo instruction bleed-through the release amplified: runtimes that auto-load directory-scoped instruction files picked up templates/CLAUDE.md, and because @-imports resolve relative to the containing file, its live import block pulled the placeholder template stubs (templates/AGENTS.md, templates/memory/* — {{PROJECT_NAME}} junk, "last_session: (none yet)", conflicting "Identify yourself as…" lines) into context as if they were real instructions. Tool-repo-only: installed targets have no templates/ directory.

Fixed

• templates/CLAUDE.md + templates/GEMINI.md no longer carry live import lines. The block is now a {{BOOTSTRAP_IMPORTS}} placeholder (the established {{UPPER_SNAKE_CASE}} mechanism); ENABLE.md Step 6 defines each vendor's literal block (Claude @path idiom / Gemini @./path.md) and expands it at install. Installed output is byte-identical to v4.29.0's — enabled repos need nothing beyond a version stamp (see the 4.29.0 → 4.29.1 rung).
• The 4.28.4 → 4.29.0 rung no longer points at the templates for the block — it references the literal blocks in ENABLE.md Step 6, so a ladder run can't copy the placeholder verbatim.

Changed

• Honest residual, documented: a runtime that auto-loads a nested AGENTS.md directly may still surface templates/AGENTS.md itself — behavior that predates v4.29.0 and belongs to the runtime; this patch removes the amplification (the memory-stub pull-in via imports), which is the part the tool controls. Root CLAUDE.md/GEMINI.md keep their live imports — that's the feature, dogfooded.
• Lockstep: VERSION → 4.29.1; CHANGELOG; README row; UPGRADE.md row + rung (+ the 4.29.0 rung amendment); templates/CLAUDE.md/GEMINI.md; ENABLE.md Step 6. No memory-file shape change; skills/adapters unchanged.

Version 4.29.0, 7/12/2026

Before-session context presence — bootstrap @-imports + an opt-in SessionStart recipe (MINOR). A child-repo field report (2026-07-11) showed agents skipping the before-session read chain (CLAUDE.md → AGENTS.md → memory/*) under task pressure — skill-unawareness, off-model engagement, rework, apparent "memory loss." The root cause is structural, not a one-off lapse: the v4.19.0 trigger layer reinforces only the after-session rituals, because its vendor-neutral substrate (git + CI) has no session-start moment — so the before-session read rested on prompt adherence, the same non-determinism v4.20.1 recorded for Copilot self-init. This closes the presence half of that asymmetry; attendance remains agent judgment.

Added

• Native @-imports in the bootstrap pointers. templates/CLAUDE.md and templates/GEMINI.md (and this repo's own root pointers — dogfooded) now import @AGENTS.md, @memory/instructions.md, @memory/continuity.md, @memory/vision.md, so the hub + core memory files are structurally present at session start on import-capable runtimes (Claude Code: @path, relative to the containing file, max 4 hops, code blocks skipped; Gemini CLI: the @./path.md form, .md files only, max depth 5). Markdown-only, no hooks — the same fix-shape as v4.20.1's copilot-instructions front-load. Imports live only in the per-vendor bootstrap files; AGENTS.md stays vendor-neutral.
• Opt-in SessionStart injection recipe in docs/optional-ritual-hook.md (tool-only, never installed) — "Option A0," for teams that also want memory/sessions/ recency injected (imports can't express dynamic paths). Deliberately not installed by ENABLE.md: the installed .gitignore ignores all of .claude/ (personal runtime config; a committed settings.json accretes personal allow-lists — observed in the field), and the hook is Claude-Code-only — Layer-2 in the trigger-design taxonomy, a per-vendor nicety, not the vendor-neutral floor.

Changed

• Honest limits, stated where they bite: the bootstrap note tells the agent imports can't express dynamic paths — the newest 2–3 memory/sessions/ logs stay an explicit read; Cursor/Windsurf/Copilot have no import mechanism and keep the prose pointer (Copilot's mitigation remains the v4.20.1 front-load pattern). The imported files enter context every session, so the continuity-bloat controls (v4.24.0/4.28.2/4.28.3) are now load-bearing, not cosmetic. Attestation canaries (a child repo's local hardening) remain a downstream per-repo pattern, not part of the tool.
• Lockstep: VERSION → 4.29.0; CHANGELOG; README row; UPGRADE.md row + 4.28.4 → 4.29.0 rung; root + template CLAUDE.md/GEMINI.md; docs/optional-ritual-hook.md (retitled "Reinforcing the session rituals with hooks" — it now covers a before-session layer). No memory-file shape change; skills/adapters unchanged.

Version 4.28.4, 7/6/2026

Co-Authored-By trailer duplication — reframed as a dedup-by-email invariant (PATCH). Third field report from mercury-composable: the v4.28.0 commit-attribution guidance assumed the agent fully authors the message, but the agent co-authors it with its harness, which often injects its own (model-version) Co-Authored-By. A conscientious agent then appended a second (stable-name) trailer → duplicate co-authors for one collaborator; squash-merges compounded it (one observed commit had 55 trailer lines). Doc-only refinement of the convention; no code, no shape change.

Changed

• The rule is now a deterministic invariant, keyed on email. AGENTS.md (root + templates/) reframes the mental model — you co-author the commit message with your harness; treat the harness's message as the base and reconcile, don't blindly append — and states the invariant: at most one Co-Authored-By per collaborator, matched on email (Claude Code, Claude Opus 4.8, Gemini CLI, Gemini 2.5 … are one collaborator at one noreply@… address; the name varies, the email doesn't). A three-branch resolution tree replaces the ambiguous "accept it": harness trailer you can't suppress → that's your one; you control the message → emit exactly one stable-name trailer and drop the harness duplicate at the same email; never both.
• Squash-merge guidance is forge-aware. The canonical trailer lives once (PR-description footer; omitted from per-commit bodies); trim inline repeats so one line per collaborator survives.
• PR template + docs. The .github/pull_request_template.md footer comment (root + templates/) and docs/getting-started.md now carry the one-per-collaborator-by-email rule.
• Enforcement deferred (deliberate). A dedup hook and a memory-lint advisory were proposed; both deferred — an auto-dedup commit-msg hook would rewrite commit messages, a departure from the tool's "never mutate your commits" stance. Revisit (an observational lint advisory first) if the doc fix doesn't hold in the field.
• Lockstep: VERSION → 4.28.4; README row; UPGRADE.md row + 4.28.3 → 4.28.4 rung; docs site. Re-sync AGENTS.md + .github/pull_request_template.md on enabled repos. No memory-file shape change; skills/adapters unchanged.

Version 4.28.3, 7/6/2026

[continuity-bloat] line-count message is now decay-aware (PATCH). Second field report from mercury-composable (a 29-module Maven reactor): after a by-the-book review the fact count is healthy but the continuity_max_lines backstop trips on genuinely-active, information-dense Key Decisions — and the review has nothing to archive, so the warning can't be honestly cleared. Same failure class as v4.28.2 (an unclearable warning erodes the signal), but on the line axis; the generic "a review is due to lean it down" also nudges toward premature archival of active facts — REVIEW.md's costliest error.

Fixed

• The line-count [continuity-bloat] message branches on whether anything is actually archivable. memory-lint now computes archivable = facts overdue for decay (sslu > archive_window, excluding core/superseded/pinned) + superseded facts. When lines exceed continuity_max_lines and archivable == 0, the message tells the truth — "nothing is archivable yet; the excess is active/dense facts. Condense shipped decisions, or raise continuity_max_lines in decay-policy.md if this repo is legitimately large." — instead of prescribing a review that can't help. When something is archivable, the original actionable message stands. Validated on mercury's live memory (37 facts / 708 lines / 0 archivable → the honest message; both runtimes byte-identical).
• decay-policy.md note. The continuity_max_lines comment now states it is meant to be raised for a legitimately large/complex repo (covers the field report's per-repo-acknowledgement ask cheaply; the value was already per-repo tunable).
• Scope / deferred. Only the line-count message changed; the fact-count check (v4.28.2) is untouched. A dedicated non-archival "condense shipped decisions" lever was proposed but deferred until the need recurs in the field — the honest message points at that lever without building it yet.
• Lockstep: VERSION → 4.28.3; check_continuity_health gains an optional archivable arg (default unknown → generic message; backward compatible) in both runtimes at parity; main() computes and passes it; two new mirror tests (archivable > 0 → actionable, archivable == 0 → active-verbosity). README row; UPGRADE.md row + 4.28.2 → 4.28.3 rung. Skill description unchanged → adapters need no re-sync.

Version 4.28.2, 7/4/2026

Fix [continuity-bloat] false positive on mature repos (PATCH). Field report from mercury-composable: the bloat check fired permanently after a correct review because it counted tier: core facts and pinned open threads against the cap, even though those entries can never be archived. The fix aligns the count with decay-policy.md's documented intent ("count of decaying facts/threads") by excluding non-archivable entries.

Fixed

• memory-lint bloat check now counts only decay-eligible facts. tier: core entries (structural invariants) and pinned open - [ ] threads are excluded from nfacts before comparing against continuity_max_facts. A repo with 14 core facts + 11 open threads + 16 working facts now correctly reads 16 < 30 (clean) rather than 41 > 30 (chronic WARN) after a fully correct review. The warning message updated to say "decay-eligible facts" so the scope is unambiguous.
• Lockstep: VERSION → 4.28.2; check_continuity_health gains optional pinned param (default set() — backward compatible); call site passes the already-computed pinned set; new test test_fact_bloat_excludes_core_and_pinned covers the mercury scenario directly.

Version 4.28.1, 7/2/2026

Post-commit hook: uncommitted-session-log guard (PATCH). The auto-stub window check misfired on the recommended two-commit pattern (feature commit → chore(memory) commit) when the session log was written more than 30 min before the feature commit — the clock is the wrong proxy when the artifact state is directly observable.

Fixed

• Suppress the auto-stub when a session log already exists uncommitted. The .githooks/post-commit auto-stub gained a primary guard: before the filename-timestamp window check, it inspects git status --porcelain -- memory/sessions/ for any uncommitted .md. If one is staged, modified, or untracked, the agent already has a log for this session — the hook emits the enrich-and-commit nudge and skips stubbing. Previously the filename-timestamp threshold treated an in-flight (uncommitted) log written >30 min before the feature commit as "too old" and stubbed a near-duplicate. The window check stays as fallback for already-committed, hours-old logs (e.g. a follow-on commit). Bash 3.2-compatible; hook stays non-blocking.
• Lockstep: VERSION → 4.28.1; UPGRADE.md row + 4.28.0 → 4.28.1 rung. Re-copy .githooks/post-commit. No memory-file shape change; no skill/adapter change.

Version 4.28.0, 6/29/2026

Co-author convention cleanup — stable agent identity + one trailer (MINOR). From a real-world finding dogfooding the v4.27.0 convention on mercury-composable: a squash-merged PR (#126) surfaced two imperfections in how the Co-Authored-By self-identification renders. The attribution itself worked (human + AI both credited, Verified) — these are consistency fixes.

Changed

• Stable agent identity, not a model version. The convention said "identify yourself the same way you do in session logs," but session logs use a stable agent name (Claude Code) while the commit/PR trailer was a model-version string (Claude Opus 4.8) — so it (a) wasn't actually "the same," and (b) churns every model release, fragmenting the co-author across many identities over time. AGENTS.md (root + templates/) and the PR-template footer comment now specify the stable agent name — the actual AI collaborator (Claude Code, Gemini CLI, …), kept vendor-neutral via a <your agent name> placeholder — explicitly not a model version. Where a runtime auto-injects a model-version commit trailer it can't override, that's accepted (the agent controls the PR footer + session logs, which carry the stable name).
• One trailer on squash-merge. Every commit carries the trailer (often harness-injected), so a squash-merged PR stacked ~9 identical inline Co-Authored-By lines plus the consolidated one GitHub appends after the --------- separator. The convention now notes: on a squash-merge, collapse to a single Co-Authored-By: — GitHub's appended consolidated line is canonical; trim the redundant inline repeats. Advisory (the agent can't suppress the harness trailer or GitHub's squash template; this is merge-time guidance).
• Lockstep: VERSION → 4.28.0; UPGRADE.md row + 4.27.0 → 4.28.0 rung; README row (and the recent-releases table trimmed back to its 10-most-recent cap); docs site (Getting Started). No memory-file shape change; no skill/adapter change.

Version 4.27.0, 6/29/2026

Standardized PR description — lead with What / Why (MINOR). From a maintainer suggestion: standardize the look-and-feel of pull-request descriptions created under the agent-memory protocol with two short summary sections — What and Why — so the standard propagates to every AI-enabled repo. A natural fit: why is already a first-class artifact throughout the protocol (the VBDI intent trace, ADR rationale, supersession reasons, even this changelog's own > summary + Why shape), and the content projects straight from the session log(s) in the PR — What from the change, Why from the Blueprint gap / decision it serves. There was no PR guidance before this (only a commit Co-Authored-By trailer convention), so it fills a real gap.

Added

• .github/pull_request_template.md (this repo + templates/, so every enabled repo inherits it) — two sections, What (the change) and Why (the intent it serves; substantive intent, not a restatement of What), each 1–2 short paragraphs (flexible — a trivial PR isn't padded; a What may use a bullet or two). GitHub auto-populates the PR body from it; the agent draws the content from the session log(s) in the PR. The template also closes with a self-identifying Co-Authored-By: footer — the same authorship-traceability convention as commits and session logs, now extended to the PR altitude. Advisory, never a gate (guide-don't-prescribe). Installed by ENABLE.md Step 6 (tracked — it travels).

Changed

• AGENTS.md (root + templates/) — the after-session commit step gains an "Opening a pull request? → lead with What / Why" convention next to the co-author-trailer note, plus a checklist line. This is the vendor-neutral backstop: the .github/ template only covers the GitHub web UI / gh pr create, so the steering line carries the convention for agents that compose a PR body themselves. Re-sync a target's AGENTS.md from templates/AGENTS.md.
• ENABLE.md Step 6 installs the new template; UPGRADE.md gains the 4.26.1 → 4.27.0 rung + version-table row; README version-table row + file-tree entry; docs site updated. VERSION → 4.27.0. No memory-file shape change.

Version 4.26.1, 6/28/2026

[stale-metadata] / refresh-metadata no longer opine on a pinned thread's tier (PATCH). A sanity check of mercury (post-Copilot review) surfaced that v4.26.0 flagged every working-tagged pinned - [ ] open thread as "should be active" — noise, since a pinned thread never decays regardless of its tier label. The protection is its pinned-ness (being unchecked), not the label. Refinement: the tooling now has no opinion on a pinned thread's tier — memory-lint won't flag it and refresh-metadata won't rewrite it (it still refreshes the factual uses/last_used). Surfaced comparing refresh-metadata against Copilot's own update-metadata.py (which skipped pinned threads); this lands on the same outcome, by a cleaner rule.

Changed

• memory-lint expected_tier (check 9) + refresh-metadata expected_tier: a pinned thread returns its stored tier (no drift, no rewrite) instead of being forced to active. Both runtimes at parity; tests added (memory-lint 34, refresh-metadata unchanged count — assertion updated).
• DECAY.md rule 4 + both SKILL.md notes clarify: pinned-ness protects an open thread, not its tier label; working on a pinned thread is fine. Doc-only otherwise; no shape change.

Version 4.26.0, 6/28/2026

refresh-metadata + a memory-lint [stale-metadata] advisory — close the skipped-re-tier gap (MINOR). From a cross-vendor field test: Copilot / Gemini 3.1 Pro committed the v4.25.0 upgrade to mercury and, seeing the [review-overdue] advisory, ran the review unprompted — correctly using archive-fact (the over-archival guard even caught a premature-archive attempt and it reverted). But it did Step 4 (archive) + Step 5 (sweep) and skipped Step 2 (apply events) + Step 3 (re-tier) — leaving stale last_used/uses/tier footers on the facts that stayed. A capable agent silently did half the ritual. This is the third instance of one failure class (after the truncation trap → archive-fact, and the never-fired review → [review-overdue]): multi-step agent rituals get partially executed. The fix refines the earlier "reject auto-fix" stance — the real boundary is judgment vs. arithmetic: deciding what to archive is judgment (stays with the agent); recomputing metadata is arithmetic (the "full rebuild" path REVIEW.md already calls deterministic + reproducible) and is safe to mechanize.

Added

• agent-skills/refresh-metadata/ — the 7th built-in (provenance: agent-memory-builtin). Executes REVIEW.md steps 2–3 deterministically: recomputes every fact's last_used / uses / tier from the ## Memory References across memory/sessions/ and writes the footers back (reads into memory, writes once). Pure arithmetic — core/superseded and never-referenced facts are untouched, and it clamps at archive-candidate (it never archives; that's archive-fact + the agent's judgment). Python or Node at parity, with mirror tests; --dry-run; idempotent.
• memory-lint check (9) [stale-metadata] — flags a fact whose stored tier ≠ the tier recomputed from the reference log (i.e. steps 2–3 were skipped), excluding core/superseded/never-referenced. Both runtimes + mirror tests. This is what makes the skipped-re-tier gap visible on every lint run + CI.

Changed

• REVIEW.md steps 2–3 now lead with refresh-metadata (the deterministic path; hand-editing 30 footers is discouraged). ENABLE.md §5i installs seven built-ins. README / ADR / continuity built-in lists updated. sync skill adapters now writes 8 skills → 48 adapters.

Version 4.25.0, 6/28/2026

archive-fact — a deterministic, safe archive-move helper (MINOR). From a cross-vendor critique (Copilot / Gemini 3.1 Pro): "agent behaviors vary by vendor; the reliance on agent interpretation of REVIEW.md to safely mutate state is vulnerable to LLM regressions / file-editing precision — harden the memory-writing mechanism itself, perhaps a small CLI helper for safe writes." Spot-on, and it names our most-repeated bug: the open(f,"w").write(open(f).read()+…) truncate-before-read trap wiped a version.md stamp and then this repo's archive (50 facts → 6), once each. The v4.22.4 safeguard moved the rule from a personal note → shared doc; this is the logical next step — doc → tool.

Added

• agent-skills/archive-fact/ — the 6th built-in (provenance: agent-memory-builtin). Executes REVIEW.md step 4's archive-move deterministically: given fact id(s) the agent has decided to retire, it extracts each block (the column-0 bullet through its <!-- id: … --> footer, whole), appends it to the quarter archive + INDEX.md (append-mode), and rewrites continuity.md without it (read whole file into memory, write once — truncation is structurally impossible). Python or Node at output parity, with mirror tests (test_archive_fact.py / .mjs). Guards (refuse, exit 1): id with no footer, id already archived, or a move that would empty continuity; the move is all-or-nothing. --dry-run to preview. Keeps the meaning/mechanics split intact (same as memory-lint): the agent judges what to archive; the helper performs the move safely — it never decides (never-pick-a-winner).

Changed

• REVIEW.md step 4 now leads with archive-fact as the preferred path (manual append-mode / read-into-var stays the no-runtime fallback). ENABLE.md §5i installs six built-ins. README.md / ADR.md / continuity.md built-in lists updated. sync skill adapters now writes 7 skills → 42 adapters.

Critique disposition (also from the same review)

• Auto-fix, not just lint — declined for semantic decay (auto-deciding what to archive violates never-pick-a-winner); the REVIEW.md ritual is the agent-driven fix, the v4.24.0 advisory triggers it, and archive-fact now makes the move deterministic. Upgrade automation — logged as a backlog thread (the mechanical steps could be scripted, but the rungs are semantic merges that must stay agent-judged; the chosen mitigation is deterministic guardrails like memory-lint).

Version 4.24.0, 6/28/2026

Decay-policy retune + a review-cadence/size advisory in memory-lint (MINOR). Grounded in real measurements across two enabled repos: this tool (121 sessions, continuity 490 lines / 24 facts right after a clean review) and a product repo (61 sessions, 585 lines / 41 facts, 0 archived — the cadence review never fired in the field). Findings: (1) continuity_max_lines: 300 was permanently-red on both (alert fatigue) and conflates non-decaying structural sections (Vision, Invariants, Key Decisions, Blueprint) with the decaying region; (2) there was no count-based trigger, though fact-count is the verbosity- and velocity-independent signal; (3) nothing enforced the review cadence, so a real repo silently accumulated; (4) verify_invariants_every: 20 meant near-daily human re-confirms at burst velocity (10–20 sessions/day).

Changed

• memory-lint — new advisory check (8), both runtimes at parity + mirror tests (29 each). [review-overdue] fires when sessions_since_last_review ≥ review_every (read from the last_review Project-State stamp; never reviewed ⇒ all sessions count). [continuity-bloat] fires when continuity holds more than continuity_max_facts decaying facts/threads, or more than continuity_max_lines lines. All advisory (WARN) — a review is a ritual, never a hard gate — but they ride every lint run + the CI floor, so a lapsed review can't hide. This is the keystone: parameters alone don't fix a review that never runs.
• Decay-policy defaults retuned (templates/memory/decay-policy.md + this repo's): continuity_max_facts: 30 (NEW — primary lean signal), continuity_max_lines: 300 → 600 (a mature layer sits ~450–600 when healthy), verify_invariants_every: 20 → 40 (anti-fatigue). working/active/archive_window and review_every unchanged — they work; the bloat came from reviews not running, not from the windows.
• Docs: REVIEW.md size trigger (facts primary + the new lint advisories), agent-skills/memory-lint/SKILL.md (check 8), AGENTS.md (root + template) review-cadence note, templates/.agent/schema.md windows list. Skill description unchanged → adapters need no re-sync.

Version 4.23.2, 6/28/2026

Context-hygiene guidance: keep state externalized so compaction is safe (PATCH). From a maintainer exchange about multi-hour AI sessions. Two corrections to an initial "brain fog" framing: (1) the agent usually can't compact itself (compaction is user-invoked or a harness auto-compact), so the agent's reliable lever is externalizing state, not self-compacting; (2) "retrieval degrades over hours" is a judgment call — the objective health signal is context-window utilization (tokens used vs. the model's limit), which the harness tracks and may auto-act on, while time and a felt "fog" are only proxies. AGENTS.md now teaches: write the session log + continuity.md at each natural seam (milestone, phase shift, before a topic pivot) before compaction; at a seam with high utilization, suggest compacting (or rely on auto-compact), never mid-task; re-verify specifics against live files after any compaction. Reinforces why the memory layer lives in files, not the chat buffer.

Changed

• templates/AGENTS.md + root AGENTS.md — new "Long session? Keep state externalized so compaction is safe" block, placed as a corollary to lightweight mode (the session-log write is the seam). Vendor-neutral (/compact / auto-compact / fresh session). Doc-only, additive, no memory-file shape change; re-sync templates/AGENTS.md into targets on the 4.23.1 → 4.23.2 rung.

Version 4.23.1, 6/28/2026

last_harvest marker for incremental harvests (PATCH). From a cross-vendor test drive: mercury's agent ran harvest-knowledge correctly but had to infer the harvest window ("since enable") because nothing recorded when the last harvest ran. Add a last_harvest marker so each run scopes "docs changed since then" deterministically. Decided to keep it in continuity.md Project State (with last_review / last_invariant_check — same family: "when did this periodic memory ritual last run"), not .agent/version.md (the install manifest, a different concern).

Changed

• templates/.agent/schema.md — Project State gains an optional last_harvest: YYYY-MM-DD | through <session-file> field (omitted until the first harvest).
• agent-skills/harvest-knowledge/SKILL.md — now reads last_harvest to scope the next run to docs changed since then (full first pass if absent), and stamps it on completion (even on a no-op run — "docs checked through here"). The harvest owns last_harvest as the review owns last_review. The check-existing-first guard still prevents duplicates, so a re-scan is always safe — last_harvest only scopes the read. VERSION → 4.23.1; re-synced on the 4.23.0 → 4.23.1 rung (skill description unchanged → adapters unchanged).

Version 4.23.0, 6/28/2026

harvest-knowledge — on-demand knowledge harvest as a built-in skill (MINOR). The curious harvest (v4.22.0, ENABLE.md Step 4b) seeds memory from the team's docs once, at enable. But a living repo's docs keep evolving, and there was no recurring way to fold new ADRs / design specs / decision-log entries into memory short of re-enabling. This adds a 5th built-in skill that does exactly that, on demand — and moves "re-harvest" out of the Mode B upgrade path (it was a one-time backfill bolted onto a rung) into a proper, repeatable skill. The enable-time harvest stays a fresh-enable event.

Added

• agent-skills/harvest-knowledge/ — a no-code, agent-run built-in (provenance: agent-memory-builtin). Re-scans the repo's human-authored documentation (recursive docs/ trees + root sweep for ADRs, decision logs, design specs, roadmaps, kanban — same net as Step 4b) and distills durable facts into the neutral, shared memory/ layer additively: map-don't-mirror, check-existing-first so a re-run doesn't duplicate, conflicts → a Contradiction thread (never-pick-a-winner), genuine replacements → supersession, budget-with-disclosure → a (knowledge-harvest) Open Thread. Installed into every repo by ENABLE.md §5i (now five built-ins); added to the upgrades-additive managed-built-ins list. Explicitly not a vendor /init: /init does a deep code analysis and (re)writes/overwrites a vendor steering file; harvest-knowledge reads human-authored knowledge, writes neutral shared memory, and is additive + repeatable — borrow /init's analysis muscle, but the output goes to memory/, never a vendor file.

Changed

• UPGRADE.md — the 4.21.0 → 4.22.0 rung no longer offers an inline Mode B re-harvest; backfilling an already-enabled repo from docs is now the harvest-knowledge skill's job (run on demand). The enable-time harvest remains a fresh-enable event. VERSION → 4.23.0; the 4.22.4 → 4.23.0 rung installs the new built-in.

Version 4.22.4, 6/28/2026

Safe-write safeguard in REVIEW.md (PATCH). A truncate-before-read antipattern (open(f,"w").write(open(f).read() + …)) wiped this repo's archive during a review (50 facts → 6; caught by memory-lint, recovered from git). The lesson belonged in the shared, committed protocol (every contributor + every vendor's agent inherits it), not a per-machine personal note — so it's now a rule in the review ritual's Safety section.

Changed

• REVIEW.md → Safety gains two rules: (1) never truncate a memory file when scripting the move — append with >> / append-mode for the archive + INDEX.md, read-into-a-variable-then-write for continuity.md; never open(f,"w").write(open(f).read() + …) (opening "w" truncates before the read). (2) After any scripted memory mutation, run memory-lint — it catches a truncation (count drops, links dangle), and git-tracked memory files recover via git checkout HEAD -- <file>. Installed verbatim into every enabled repo; re-synced on the 4.22.3 → 4.22.4 rung. VERSION → 4.22.4.

Version 4.22.3, 6/27/2026

Tighten the post-commit session window: 2h → 30 min (PATCH). v4.22.1 shipped a 2-hour active-session window, but the observed problem was follow-up stubs minutes apart, and 2h is long enough to wrongly conflate a genuinely new session that starts within 2h of the previous one's log (the hook would nudge "enrich the old log" for a new session). 30 minutes spans a working session's commit cadence — including a short test/think gap — while a real new session (a longer pause) still gets its own stub.

Changed

• .githooks/post-commit — the active-session window default is now 30 minutes (was 2h), and the override env var is AGENT_MEMORY_SESSION_WINDOW_MINUTES (integer minutes), replacing AGENT_MEMORY_SESSION_WINDOW_HOURS. The unit changed to minutes because BSD date -v rejects a fractional hour (-v-0.5H), so sub-hour windows need integer minutes; both runtimes stay supported (date -v-30M / date -d "30 minutes ago"). Re-validated (2 min & 25 min → suppress; 45 min & none → stub; custom override honored). .githooks/README.md + docs/DESIGN-ritual-triggers.md updated; VERSION → 4.22.3; re-copied into targets on the 4.22.2 → 4.22.3 rung.

Version 4.22.2, 6/27/2026

Lightweight mode: one log per working session, not per commit (PATCH). The agent-side mirror of v4.22.1. The lightweight-mode rule wrote a lite session log for every memory-neutral tracked-file change — so an agent making several small commits in one sitting produced a near-duplicate lite log per commit (the same clutter + decay session-count inflation the v4.22.1 hook fix addressed, but on the agent-behavior side). The agent has context the hook lacks — it knows it's the same working session — so it needs no time-window: it simply enriches the log it already wrote.

Changed

• AGENTS.md (root + templates/) — the lightweight-mode "lite log" tier gains: if you already wrote a session log earlier in this working session, a later memory-neutral commit should enrich that existing log (a one-line "also: …" note) rather than spawn another lite log — a burst of commits in one sitting is one session. Explicitly preserves the existing model for memory-relevant work (distinct events still each get their own log, so a multi-task conversation may still yield several); the rule only stops trivial follow-on commits from minting near-duplicate lite logs that would also inflate the decay session-count. Wording-only — no file shape, skill, or script change. VERSION → 4.22.2; re-synced into targets on the 4.22.1 → 4.22.2 rung (re-copy templates/AGENTS.md).

Version 4.22.1, 6/27/2026

post-commit auto-stub is now per session, not per commit (PATCH). A downstream report (mercury-composable) found the hook's auto-stub fired on every work commit: its only de-dup guard checked for an untracked stub, so once the agent enriched and committed the session log, the next work commit found no waiting stub and wrote a fresh one. One long session produced ~6 near-identical lite logs (each with an extra memory: session log … commit), and — since the decay model counts session files — inflated sessions_since_last_used, so facts decayed ~N× too fast (ironic, given memory-lint exists to catch decay miscounts).

Fixed

• .githooks/post-commit — the auto-stub now suppresses a new stub when a session log already exists within an active-session window (default 2h; override AGENT_MEMORY_SESSION_WINDOW_HOURS), nudging the agent to enrich that log instead. Detection is by the newest session filename timestamp (immutable + clone-safe — mtime is reset by git clone/checkout, which would wrongly suppress stubs for hours after a clone), compared lexicographically against a window-ago stamp in the same YYYY-MM-DD-HHMMSS form. This subsumes the old untracked-stub guard (a fresh stub has a recent filename) and covers the just-committed case too. A genuinely new session (no log within the window) still gets a stub — the "no silent gap" guarantee holds. Net: one enriched log per working session, and the session-file count tracks sessions, not commits. Bash 3.2-compatible (BSD date -v with a GNU date -d fallback). VERSION → 4.22.1; re-copied into targets on the 4.22.0 → 4.22.1 rung.

Version 4.22.0, 6/27/2026

Discovery, consent & merge-friendliness + MERGE.md (MINOR). One release bundling four additive improvements developed iteratively in a single session (dev-numbered 4.22–4.25) and consolidated into one version per a new "one version per release" policy — VERSION and the upgrade ladder track releases, not per-feature dev iterations (granular history lives in the session logs + Open Threads). The work came from three real signals — a client enablement complaint, a teammate-concurrency review, and a GitHub Copilot review — and was hardened by a full fresh-context review loop validated end-to-end by a different vendor: Gemini 3.1 Pro critiqued the milestone and re-validated the applied fixes, proving the second-opinion → apply-critique pair cross-vendor on this very release.

Added

• Curious knowledge harvest at enable (ENABLE.md Step 4b) — discovery now recursively descends every documentation tree (docs/, doc/, wiki/, rfcs/, adr/, … all subfolders) and sweeps repo + module roots for human-authored knowledge markdown (decision logs, ADRs, kanban/roadmap, architecture & design notes, CONTRIBUTING, RFCs), distilling the durable facts into memory (map-don't-mirror). Bounded by a read budget with disclosure — overflow is recorded as a (knowledge-harvest) Open Thread so nothing vanishes silently. From a client complaint: a canonical docs/ tree was skipped entirely and a recursive re-ask still stopped at depth-1.
• Fresh-enable advisory + discovery depth (ENABLE.md Mode A) — a fresh enable now opens with a concise exec summary of the protocol (what it is, what it writes, what it leaves untouched, that it's committed + shared) and a cancel gate, for informed consent rather than a blind yes. It then offers a discovery depth: standard scan (default) or an /init-depth deep analysis — which borrows the capability of a vendor's built-in /init but writes findings into the neutral memory layer, never a vendor steering file. A first enable session log records the chosen depth.
• MERGE.md — a new installed, no-code, on-demand protocol for resolving a git merge/rebase conflict in memory/. Tiered + human-gated, enforcing the never-pick-a-winner invariant: classify the hunk → Tier-1 mechanical (additive → union/keep-both; scalar → take-later, deterministic) → Tier-2 semantic clash (the AI never decides — preserve both + raise a Contradiction; a supersession only on the human's explicit instruction) → memory-lint gate → human approves the merge commit (never auto-commit). Joins DECAY/REVIEW/SKILLS as a root protocol doc, read on demand.
• memory-lint check 7 — a leftover VCS merge-conflict marker (<<<<<<< / >>>>>>> / diff3 |||||||) in the live top-level memory/*.md files is now an ERROR (a bare ======= line is exempt — it's a valid Markdown setext heading underline). sessions/ and archive/ are excluded (immutable/append narrative that legitimately quotes markers — e.g. a session log pasting a diff). Both runtimes at parity, with mirror tests.

Changed

• continuity.md merge-friendliness — status is now spec'd as a short current-state line, not a changelog (.agent/schema.md + AGENTS.md): a single ~20 KB accreted status line was a merge-conflict hotspot for concurrent teammates and duplicated the upgrade ladder. A new schema section "Concurrency & merge-friendliness" documents the conventions (one fact per line; append-only sections → keep-both/union on conflict; scalar bumps → take-later; same-thread clash → human). This repo's own status line slimmed 19,828 → ~1,100 chars.
• "One version per release" policy added to UPGRADE.md's versioning model — bump for a release event, not per feature; keep a single pending version while unreleased; magnitude = the largest change in the batch; consolidate working-tree bumps beyond HEAD before committing. (This release is the first to apply it: four dev-iterations → one v4.22.0.)
• upgrades-additive invariant + ADR-0005 — added sync-adapters to the managed-built-ins example list (it became a tool-managed built-in in v4.18.0). Surfaced and applied during a maintainer invariant re-verification (all five core invariants + the Vision re-confirmed, none superseded).
• REVIEW.md step 9 now codifies that archived ids go in the ## Memory Review block, never under ## Memory References (which re-arms the over-archival guard → false reactivation). .githooks/README.md documents that a feat/memory commit split trips the advisory hook on the code commit (expected/benign; prefer an atomic commit). VERSION → 4.22.0.

Notes

• Mostly operator-side (ENABLE.md) plus the memory-lint skill and one new root doc (MERGE.md) — no memory-file shape change. The single 4.21.0 → 4.22.0 upgrade rung bundles all four features.
• A full memory review ran alongside this release (continuity 753 → 577 lines; 14 faded facts archived).

Version 4.21.0, 6/26/2026

Google Antigravity (agy) skills adapter (MINOR). Adds a 6th skills adapter target, .agents/skills/<name>/SKILL.md. Antigravity — the successor to Gemini CLI — merged custom commands into the open Agent Skills standard and reads workspace skills from .agents/skills/, not the old .gemini/commands/*.toml. So on an enabled repo, init.sh faithfully populated the Gemini adapter, yet agy reported /<command> (e.g. /memory-lint) as not found — it never looks in .gemini/commands. The new adapter uses the same SKILL.md shape as the Claude/Kiro/Copilot adapters, so it is a one-line addition per runtime, no new format. The .gemini/commands TOML adapter stays for now so Gemini CLI keeps working through the transition. Found dogfooding agy on an enabled repo.

Added

• 6th vendor adapter .agents/skills/<name>/SKILL.md in all three sync-adapters runtimes (sync-adapters.py / .mjs / .sh, kept byte-for-byte equivalent — verified identical output): each now writes six adapters per skill and prunes orphaned .agents/skills/ dirs by the shared pointer signature. .agents/ added to the root and template .gitignore (regenerated, local-only).
• Google Antigravity (agy) added to the Mode C detection/migration table and the bootstrap-files table in README.md (reads AGENTS.md + .agents/skills/; Gemini CLI successor). Vendor adapter lists updated across ENABLE.md, templates/AGENTS.md, templates/.agent/schema.md, and the sync-adapters SKILL.md (five → six). VERSION → 4.21.0.

Version 4.20.3, 6/24/2026

memory-lint catches an empty/malformed version manifest (PATCH). Adds a deterministic check_version_manifest check (an ERROR) so a present-but-empty or malformed .agent/version.md is caught by the lint floor (CI + reviews + manual runs) instead of silently breaking Mode B upgrade detection. This closes the loop on the v4.20.1 bug: a truncating stamp one-liner emptied a target's version.md, which made an agent misread the repo's version — now that exact failure mode fails the lint rather than going unnoticed. A missing version.md stays valid (the pre-versioning baseline) and is not flagged.

Added

• check_version_manifest in both memory-lint.py and memory-lint.mjs (byte-identical message, at parity) — wired into main's error list; with mirror tests in test_memory_lint.py / test_memory_lint.mjs (empty → error, malformed → error, valid → ok, missing → ok). SKILL.md documents the new check. Re-copied into targets on the 4.20.2 → 4.20.3 rung. VERSION → 4.20.3.

Version 4.20.2, 6/24/2026

Windows line-ending hardening (PATCH). Adds a .gitattributes pinning *.sh and .githooks/* to LF, so Git for Windows (core.autocrlf=true by default) doesn't rewrite them to CRLF on checkout — which would break bash (bad interpreter: /usr/bin/env bash^M) and silently disable the hook + init.sh. From a Copilot Windows-feasibility check: the bootstrap works via Git Bash/WSL, but without this it was luck-of-the-default. Now robust on Windows.

Added

• .gitattributes (root + templates/) — *.sh text eol=lf, .githooks/* text eol=lf. Installed / merged into targets additively (ENABLE Step 7b + a 4.20.1 → 4.20.2 rung), same discipline as the .gitignore block. VERSION → 4.20.2.

Version 4.20.1, 6/24/2026

Self-init folded into copilot-instructions.md (PATCH). A fresh-clone dogfood showed v4.20.0's self-init worked for Claude Code (which acts on AGENTS.md) but not GitHub Copilot CLI — its start is driven by copilot-instructions.md's front-loaded read list, so it loaded memory and summarized without acting on the AGENTS.md self-init step (hook stayed inactive, adapters absent). Fix: put the first-run init at the top of copilot-instructions.md (the file Copilot reliably front-loads), as the first action — run bash .githooks/init.sh if core.hooksPath is unset / adapters absent, before summarizing. Improves Copilot's odds of self-initing; bash .githooks/init.sh remains the one-command fallback and CI remains the zero-config floor.

Changed

• templates/.github/copilot-instructions.md — now leads with a first-run-init block. VERSION → 4.20.1.

Version 4.20.0, 6/24/2026

First-run init for fresh clones (MINOR). Closes the gap a Copilot fresh-clone dogfood exposed: the memory bootstrap self-initializes (the agent reads AGENTS.md/memory on its own), but a clone has the gitignored skill adapters absent and the git hook unactivated (git can't auto-run committed hooks on clone). That had meant a two-command manual setup. Now: .githooks/init.sh (one idempotent command — regenerate adapters + git config core.hooksPath .githooks) + an AGENTS.md self-init note so the agent does it on its first session. CI remains the zero-config floor regardless.

Added

• .githooks/init.sh — one-command first-run init (sync adapters + activate the hook). Idempotent; not a git-hook name, so git never auto-runs it (run bash .githooks/init.sh). Committed executable (100755).

Changed

• AGENTS.md (root + template) — first-session self-init note: run bash .githooks/init.sh if adapter dirs are empty / core.hooksPath is unset.
• .githooks/README.md — now leads with the one-command init.
• UPGRADE.md (rung + table), docs/DESIGN-ritual-triggers.md, README, VERSION → 4.20.0.

Version 4.19.0, 6/24/2026

Vendor-neutral ritual triggers (MINOR). The after-session ritual (session log, review, sync) no longer depends on the agent self-triggering — which failed in practice: client teams report it isn't followed through even with Claude, a Copilot-only team had no triggers at all, and the hook layer was opt-in, per-vendor, and didn't travel into targets. Enable now installs a committed, vendor-neutral .githooks/post-commit and a CI floor, and the agent activates the local hook — zero manual user step (the adoption constraint: any manual op is a barrier as the protocol gains traction and lands with untrained users). no-build-step-agent-run holds: git/CI invoke them in the user's env; the tool runs nothing. Design: docs/DESIGN-ritual-triggers.md.

Added

• .githooks/post-commit (bash, advisory, never blocks) — after a commit: auto-stubs a session log when the commit did real work but carried none (capture; the thoughtful summary stays the agent's job — same split as memory-lint) and re-syncs adapters when a skill changed. Activated by git config core.hooksPath .githooks (the agent runs this at enable). Plus .githooks/README.md.
• .github/workflows/agent-memory.yml — the CI floor: runs memory-lint + an advisory session-log presence check on push/PR with zero per-user setup (the untrained-user-proof layer). Advisory by default; opt-in gate via AGENT_MEMORY_STRICT=1.
• docs/DESIGN-ritual-triggers.md — the design (layered: agent-primary + git-hook/CI net + optional vendor hooks; the capture/judgment split; the no-code reconciliation).

Changed

• ENABLE.md — Step 6 installs .githooks/ + the CI workflow and activates core.hooksPath (agent-run, no user step); Notes scope + the honest git-clone-activation limit (CI backstops).
• AGENTS.md (root + template) — the ritual note is now "reinforced, not just documented" + a definition-of-done framing (a task that changed tracked files isn't done until its log exists).
• docs/optional-ritual-hook.md — reframed: the vendor-neutral git-hook + CI are now installed and agent-activated; this doc is the optional per-vendor end-of-turn hooks + git pre-commit reminder.
• UPGRADE.md (rung + table), README, VERSION → 4.19.0.

Honest limit

• Git cannot auto-run committed hooks on a fresh clone (security). So local hooks are agent-activated (no user step in the common path) and CI is the always-on, zero-config backstop.

Version 4.18.0, 6/24/2026

sync skill adapters is now a runnable script (MINOR). A new built-in sync-adapters skill ships a deterministic adapter-regeneration script (bash sync-adapters.sh + Node sync-adapters.mjs + Python sync-adapters.py, at three-way output parity; bash needs no runtime install — preferred). For each agent-skills/<name>/SKILL.md it (re)writes the five vendor adapters and prunes the orphans it generated (signature-guarded — never deletes a hand-authored vendor file). Why: dogfooding ~/sandbox/simple-proxy, GitHub Copilot CLI (Gemini 3.1 Pro) read SKILLS.md and still couldn't run sync skill adapters — it was a prose recipe with no executable, so the agent hunted for a non-existent command (node memory-lint.mjs sync skill adapters, an mcp-agent-memory npm package, npm run, grep for functions) and flailed. A real script removes the ambiguity. Consistent with no-build-step-agent-run (same category as the memory-lint script — an optional helper the agent/vendor/CI invokes; the tool runs nothing itself).

Added

• agent-skills/sync-adapters/ — the built-in sync-adapters skill: SKILL.md + bundled scripts/sync-adapters.sh (bash — no runtime install, preferred), sync-adapters.mjs (Node ≥ 18), and sync-adapters.py (Python 3 stdlib) at three-way byte-identical output parity. --dry-run previews; prune is signature-guarded (never deletes a hand-authored vendor file). Installed into every enabled repo (ENABLE §5i). The bash runtime is from a Copilot finding — the script must run in a non-Node project (no npm/npx/package.json).
• SKILLS.md documents a delete a skill operation — remove the neutral source (rm -rf agent-skills/<name>/), run the sync script (it auto-prunes the orphaned adapters, signature-guarded), and mark the continuity id superseded (DECAY §9). From Copilot feedback: the delete path was undocumented and required manual adapter cleanup. The script never deletes agent-skills/ itself — removing the committed source stays a deliberate, human-visible rm.

Changed

• sync skill adapters is now the script, not a prose recipe — SKILLS.md (the operation + "Authoring a skill"), ENABLE.md (§5h closing step + §5i "all four" built-ins), and UPGRADE.md (standing sync + new 4.17.0 → 4.18.0 rung + version table) now run it. The adapter recipe remains as the format spec the script implements. VERSION → 4.18.0.
• Built-ins are now four (memory-lint, second-opinion, apply-critique, sync-adapters) in the current-state docs; historical rungs/changelog entries keep their period-accurate "three".

Version 4.17.0, 6/24/2026

GitHub Copilot CLI skills adapter (MINOR). Adds a 5th vendor adapter target so a skill authored once in the neutral agent-skills/ layer is natively discoverable by GitHub Copilot CLI. Copilot CLI follows the open Agent Skills standard (the same SKILL.md shape as the Claude/Kiro adapter) and discovers project skills under .github/skills/<name>/SKILL.md, auto-matching by description (and also accepting an explicit /<name>). Found dogfooding ~/sandbox/simple-proxy: Copilot CLI could not find a skill that lived only in agent-skills/, because no .github/skills/ adapter had been generated. Mirrors the v4.5.0 Kiro addition. Cross-vendor validated 2026-06-24 by GitHub Copilot CLI (Gemini 3.1 Pro) in Agent mode: it authored a greeting skill, ran sync skill adapters, generated all five adapters, and recognized the .github/skills/ adapter (after a session restart — Copilot loads adapters at init) — "behaves exactly as documented."

Added

• .github/skills/<name>/SKILL.md adapter — sync skill adapters now writes five adapters (Claude / Gemini / Cursor / Kiro / Copilot). Recipe in SKILLS.md; same thin-pointer shape as the Claude/Kiro adapters.
• Copilot skills in Mode C migration — MIGRATE.md now detects and promotes Copilot project skills (.github/skills/, .agents/skills/) into agent-skills/, with adapters regenerated.
• Optional GitHub Copilot CLI sessionEnd ritual hook — docs/optional-ritual-hook.md documents a Copilot sessionEnd hook (the analog of the Claude Stop hook) that nudges the after-session ritual on a non-agentic vendor; advisory (echo, never blocks), opt-in, not installed by ENABLE.md (lives in repo-level .github/hooks/ or user-level ~/.copilot/hooks/). Addresses review point #4 (manual session upkeep) by making a skipped ritual visible, not enforced.
• Optional GitHub Copilot CLI postToolUse adapter auto-sync hook — docs/optional-ritual-hook.md documents an opt-in hook that regenerates the Copilot .github/skills/ adapter after a tool use, so authoring a skill no longer requires remembering to run sync skill adapters (from a real-vendor finding: agents don't reliably auto-sync). Convenience-scoped: Copilot adapter only, no prune, duplicates the recipe in bash (regenerate if the recipe changes), and /restart still loads it — the canonical path stays the agent-run sync skill adapters. The tool ships no sync script by design.

Changed

• .gitignore (template + this repo) — .github/skills/ is now ignored path-scoped. Unlike the other adapter dirs, .github/ is not ignored wholesale: it holds the tracked copilot-instructions.md and workflows/. Only .github/skills/ (the regenerated adapter) is ignored.
• templates/.github/copilot-instructions.md — now front-loads the explicit memory/ read list (not just a pointer to AGENTS.md), states the lightweight-mode logging rule (log when tracked files change; read-only Ask/Plan sessions correctly write no log), plus points Copilot at the skills layer. Copilot's Ask/Plan modes don't reliably chase a pointer chain, so the steering file carries the read list directly — a deliberate, Copilot-scoped reversal of the thin-pointer / single-source-read-order convention (reliability over DRY; Claude/Gemini/Kiro keep their thin pointers since they proactively load AGENTS.md). (Copilot later confirmed it reads memory/ in Ask/Plan and logs in Agent mode — fully lightweight-mode-conformant, so the steering wording avoids framing read-only no-log as a deficiency.)
• Adapter lists updated across living docs to include .github/skills/: AGENTS.md (root + template), SKILLS.md, ENABLE.md (Step 5h + the Step 8 completeness assertion now requires five adapters), UPGRADE.md (standing sync + new 4.16.1 → 4.17.0 rung + version table), .agent/schema.md, README.md, docs/DESIGN-skills-layer.md. VERSION → 4.17.0.
• SKILLS.md — sync skill adapters gains a hot-reload caveat: some runtimes load skill adapters only at startup (e.g. GitHub Copilot CLI parses .github/skills/ on init), so a freshly-synced /<name> needs a session restart / skills rescan to appear mid-session. From the Copilot validation above.
• Authoring convention strengthened (SKILLS.md + AGENTS.md root/template) — "Authoring a skill" is now an explicit 3-step action (write agent-skills/<name>/SKILL.md → run sync skill adapters → reload the runtime), because agents (incl. Copilot/Gemini) don't reliably auto-sync after writing the file. The tool can't force it mid-session (no-build-step-agent-run; auto-sync is guaranteed only at enable/upgrade per v4.12.0), so the instruction is now emphatic in both the on-demand SKILLS.md and the always-read AGENTS.md.

Version 4.16.1, 6/23/2026

Session filename drift fix (PATCH). Two independent gaps allowed agents to write date-only session filenames (2026-06-23.md) instead of timestamped ones (2026-06-23-153401.md): (1) the protocol said "Use date -u or equivalent", letting agents treat the context-injected currentDate (date-only) as equivalent; (2) memory-lint had no filename-format check, so drift was invisible until manual inspection. Fixed both: tightened protocol wording to an explicit prohibition, and added a check_session_filenames warning to both linters. Surfaced from drift in this session.

Changed

• templates/AGENTS.md — Step 1 and checklist: "always run date -u +%Y-%m-%d-%H%M%S" replaces "or equivalent"; explicit warning against using currentDate from context.
• templates/.agent/schema.md — same prohibition in the session-naming paragraph.
• memory-lint.py + memory-lint.mjs — new check_session_filenames warning (check 5): flags YYYY-MM-DD.md sessions; exported from the Node module. Registered in main() for both runtimes. Tests added to both suites (3 cases each, Python + Node).

Version 4.16.0, 6/23/2026

ADR default path aligned to industry convention (MINOR). The optional Architecture Decision Record log default moves from docs/ADR.md to docs/arch-decisions/ADR.md, matching the wider convention of placing the ledger in a named subdirectory (purpose-signalling; leaves docs/ root uncluttered). Surfaced from mercury-composable feedback — that project had already adopted docs/arch-decisions/ADR.md ahead of the default. Targets at the new path need no file move; targets at the old path rename on upgrade.

Changed

• templates/.agent/schema.md — section header updated to ## docs/arch-decisions/ADR.md and the one body reference to the path updated.
• templates/AGENTS.md — ADR path reference updated.
• AGENTS.md (root) — ADR path reference updated.
• DECAY.md §12 — Design primitive path reference updated.
• README.md — file-tree entry updated; new version-table row.
• UPGRADE.md — new 4.15.0 → 4.16.0 rung + version-table row.
• docs/arch-decisions/ADR.md — this repo's own ADR log moved from docs/ADR.md (dogfooding the new default).

Version 4.15.0, 6/22/2026

ADR log upkeep trigger (MINOR). The optional docs/ADR.md log (4.14.0) could be adopted but had no cue to evolve — the per-session ritual covered continuity-fact supersession yet never said to maintain a linked ADR. This adds the missing trigger: once the log exists, a new durable architecture decision, or the supersession/invalidation of a continuity fact carrying an (ADR-NNNN) tag, prompts a human-gated ledger update (add a newer ADR; mark the old Superseded/Deprecated, never delete; keep formalizes: ↔ (ADR-NNNN) in sync). Surfaced dogfooding mercury-composable's ADR opt-in.

Changed

• AGENTS.md (root + templates/) — the ADR paragraph gains an "If the log exists, keep it alive" maintenance/supersession trigger (propose; human-gated).
• .agent/schema.md (docs/ADR.md section) — a new "When to maintain it" paragraph spelling out the on-demand-but-kept-in-sync lifecycle.
• DECAY.md §12 — the Design primitive notes the ADR lifecycle is kept in sync with fact supersession.
• UPGRADE.md — new 4.14.1 → 4.15.0 rung (re-sync the guidance; merge into a repo-customized ADR note rather than overwrite) + version-table row.

Version 4.14.1, 6/20/2026

Re-synced AGENTS.md source clarified (PATCH). A cross-vendor dogfood (GitHub Copilot upgrading mercury-composable from v3.7.0 → v4.14.0) re-synced the target's AGENTS.md from the tool's root AGENTS.md — the operator/dual-mode dispatcher — instead of templates/AGENTS.md, leaving the target presenting itself as an enablement tool and referencing the non-installed ENABLE.md. The other synced docs were correct; only AGENTS.md has a different canonical source for a target, and the upgrade ladder didn't say so per-file.

Changed

• UPGRADE.md gains a "Source of truth for re-synced files" section — a per-file map: a target's AGENTS.md comes from templates/AGENTS.md (the memory hub); DECAY.md/REVIEW.md/SKILLS.md from the tool root; .agent/schema.md + bootstrap pointers from templates/. Explicit warning that the tool's root AGENTS.md (which references ENABLE.md/MIGRATE.md/UPGRADE.md) must never be installed into a target, plus a grep self-check ("AI-Enable Another Repository").
• New 4.14.0 → 4.14.1 rung that verifies and repairs a mis-synced AGENTS.md (re-copies from templates/AGENTS.md if the target's copy looks operator-sourced).

Notes

• UPGRADE.md is tool-operator-only — no target memory-file shape change; ENABLE.md (fresh enable) was already explicit and unaffected. Purely an upgrade-path clarity + self-heal fix.

Version 4.14.0, 6/20/2026

Optional Architecture Decision Record (ADR) log (MINOR). The VBDI Design altitude was realized by the ## Architectural Invariants / ## Key Decisions in continuity.md and the long-form docs/DESIGN-*.md, but there was no atomic, dated, status-tracked, never-deleted record of an individual architecture decision with its rationale, browsable by a human IT-governance audience. This adds the well-established ADR practice (Nygard 2011) as a lightweight, optional, human-facing ledger — kept token-efficient by living under docs/ and being read on demand, never in the per-session agent read path.

Added

• docs/ADR.md (this tool's own, dogfooded) — a single-file decision log, entries newest first, each with Status · Date · Abstract · Rationale (incl. consequences/trade-offs) and formalizes: / serves: trace links. Status lifecycle Proposed → Accepted → Superseded / Deprecated; nothing is ever deleted (mirrors the memory layer's supersession model, DECAY.md §9). Seeded with the 5 standing Architectural Invariants, plus the superseded no-code-markdown-only decision (ADR-0001…0006 — no-code-markdown-only recorded as ADR-0004, Superseded by ADR-0006).
• .agent/schema.md documents the optional docs/ADR.md section (format + lifecycle + map-don't-duplicate cross-linking + the "on-demand, not per-session" rule).

Changed

• AGENTS.md (root + template) and DECAY.md §12 note that the Design altitude may keep an optional docs/ADR.md log, explicitly not part of the Before-session read path.
• docs/DESIGN-vbdi-lifecycle.md §4 names the optional ADR log on the Design row.
• each memory/continuity.md Architectural Invariant title gains a visible (ADR-NNNN) pointer to its record (e.g. Target-repo scope only (ADR-0001)) — a human cue, explicitly not a prompt for the agent to open docs/ADR.md.

Notes

• Map, don't duplicate: continuity.md keeps the live what (with id); the ADR carries the durable why. The constraint text is never restated as competing truth.
• Not auto-installed. ENABLE/UPGRADE document the convention; no docs/ADR.md is created in a target — a team adopts one on demand. Additive → no memory-file shape change.

Version 4.13.0, 6/20/2026

Tool-provided (system) skills are marked + carry an upstream advisory (MINOR). The built-ins agent-memory installs into every repo (memory-lint, second-opinion, apply-critique) are tool-managed copies — overwritten on upgrade — but nothing told a target's AI that at edit time, so a local change could be made (and stranded) without anyone realizing it should go back upstream. This is exactly what happened with the memory-lint dangling-link fix in ~/sandbox/simple-proxy: it was a genuine fix, but the target session didn't surface it for back-port. (The existing warn-before-overwrite only fires during an upgrade, which that session never ran.)

Added

• provenance: agent-memory-builtin frontmatter marker on the three shipped built-ins, plus a one-line banner in each SKILL.md body. Any vendor's agent can now recognize a system skill by reading the file — no out-of-band knowledge needed. (hello-world is a dogfood sample, never installed into targets, so it is not marked.)
• Edit-time advisory (SKILLS.md → new "Tool-provided (system) skills" section): before editing a skill, check its frontmatter; if it is provenance: agent-memory-builtin, don't edit in place — advise the human and either fork under a new name (local variant) or upstream a genuine fix to the agent-memory project for back-port + validation. Production: file an issue in the agent-memory repo. Pre-release: an advisory message to the maintainer (best effort until there's an issue tracker). The upstream pointer is kept generic (not a hard-coded URL) until the project's enterprise-GitHub move.
• Upgrade-time backstop extended: ENABLE.md §5i's warn-before-overwrite now also advises upstreaming a locally-modified built-in (not just keep-yours/take-update).

Changed

• .agent/schema.md documents the optional provenance field; AGENTS.md (root + template) gains a one-line pointer; docs/DESIGN-skills-layer.md gains a status entry. DECAY.md / REVIEW.md and the adapter recipe are unchanged (adapters mirror only name + description, so the marker doesn't touch them). Lockstep: VERSION → 4.13.0, UPGRADE.md rung + table, README.

Version 4.12.1, 6/20/2026

memory-lint dangling-link check resolves supersession targets across memory/*.md (PATCH). check_dangling resolved superseded-by / supersedes links against footers from continuity.md + the archive only — so a fact superseded by one whose footer lives in another memory file (notably vision.md) was falsely flagged [dangling] … which has no footer anywhere. Found dogfooding on ~/sandbox/simple-proxy (a retired vision was superseded by a vision.md fact); the maintainer fixed it there, and this ports the fix back to the tool with a regression test.

Fixed

• load_repo now also pools footers from other memory/*.md files (e.g. vision.md), excluding continuity.md + decay-policy.md, into an extra set used only for supersession-link resolution in check_dangling — never counted as continuity/archive facts (counts, overdue, and over-archived checks are unchanged). Fixed identically in both memory-lint.py and memory-lint.mjs (code-point sort preserved → byte-identical output).
• Regression test added to both suites (test_memory_lint.py / .mjs): a continuity fact superseded by a vision.md fact must not warn, while a genuinely missing target still warns. Exercises load_repo end-to-end against a temp memory/ layer (the bug site), not check_dangling alone. To enable it, memory-lint.mjs now also exports load_repo + check_dangling (the Python module already exposed them) — additive, no behavior change.
• Lockstep: VERSION → 4.12.1, UPGRADE.md rung + table, README. Verified: both suites 8/8; memory-lint parity byte-identical on the tool's own memory (0 errors, 0 warnings).

Version 4.12.0, 6/19/2026

Enforced skill-adapter sync at enable + upgrade (MINOR). Skill adapters are gitignored per-machine pointers, so they don't travel with a clone/pull, and a rung that adds a new adapter target (e.g. Kiro in 4.5.0) left older skills' adapters incomplete. Previously enable/upgrade only did a read-only "recommend, don't run" check, so after an upgrade a skill's vendor-native adapters could be missing — blocking subsequent work that relies on native auto-trigger until the user manually ran sync. Surfaced upgrading ~/sandbox/simple-proxy (v4.4.0 → v4.11.1): the pre-existing hello-world skill was left without its Kiro adapter.

Changed

• ENABLE and every Mode B re-enable (upgrade or already-up-to-date) now run sync skill adapters as their closing skills step, instead of recommending it. For each agent-skills/<name>/ it (re)writes the four vendor adapters and prunes orphaned generated adapters. Safe to run unconditionally: it is idempotent and writes only gitignored files (never agent-skills/, never a committed file) — so it is not a version change and needs no session log. no-build-step-agent-run holds: the agent runs it during a human-invoked enable/upgrade, not a daemon or per-session automation. The per-session path still never touches skills; content-drift realignment is still the on-demand skill sanity check.
• ENABLE.md Step 8 now asserts adapter completeness (every skill has all four adapters, no orphans) — enforcement is checked, not convention.
• Touched: UPGRADE.md (standing "Skills adapter sync" section + 4.11.1 → 4.12.0 rung + version table; stale back-references in older rungs updated), SKILLS.md, ENABLE.md (Step 5h + Step 8), docs/DESIGN-skills-layer.md, VERSION → 4.12.0, README. AGENTS.md / DECAY.md / REVIEW.md unchanged.

Version 4.11.1, 6/19/2026

Review step-6 archival guard hardened against prose (PATCH). The REVIEW.md step-6 archival-verify told the agent to grep recent session files for an about-to-be-archived id and keep it if found — but a raw full-text grep also matches a prior review summary that merely names the fact while recording its decay status. Because every review that defers a fact re-names it, the guard re-armed each cycle: an archival livelock (tracked as ot-review-step6-prose, same class as the v4.10.1 prose-vs-heading bug). Found during the 2026-06-19 review.

Fixed

• REVIEW.md step 6 now defines a "use" as a ## Memory References entry, not a prose mention. It makes memory-lint the preferred archival check (it counts Memory-References only, so it is immune to the trap), and scopes the by-hand fallback to hits inside a ## Memory References block — explicitly ignoring ## Memory Review / ## What happened mentions. The verifier script itself needed no change (memref_ids was already line-anchored since v4.10.1); added memref_ids regression tests (prose/review-summary mention is not counted; block bounded at the next heading) to both test_memory_lint.py and test_memory_lint.mjs.

---

Version 4.11.0, 6/19/2026

memory-lint gains a Node runtime (MINOR). The deterministic decay-integrity check is the answer to "LLM hand-counting isn't reliable" — but that guarantee shouldn't depend on the machine having Python. memory-lint now ships in both Python and Node; on a node-only box you still get the script, not a hand count.

Added

• memory-lint.mjs — a Node port of the verifier (Node ≥ 18, built-in modules only, no npm install), at feature + output parity with memory-lint.py. SKILL.md documents both commands as interchangeable; the agent runs whichever runtime the machine has (no dispatcher, no installer — the agent is the runtime). Verified byte-identical output (default and --strict) on this repo.
• test_memory_lint.mjs — the cross-runtime contract. Mirrors test_memory_lint.py's fixtures exactly; both suites must pass, which is what holds the two implementations equivalent over time.

---

Version 4.10.4, 6/18/2026

memory-lint nested list fix (PATCH). Hardened the verifier script to handle deeply-nested lists correctly.

Fixed

• memory-lint pinned Open Threads: Fixed a bug where pinned_open_threads() incorrectly reset its state on any standard list bullet, causing it to lose track of the "pinned" state if a parent Open Thread contained deeply nested sub-items. It now correctly checks indentation level. (tracked via ot-memlint-pinned-nested.)

---

---

Version 4.10.3, 6/18/2026

Lightweight-mode wording fix (PATCH). A loose end surfaced while running sync skill adapters: the sync (re)wrote 16 gitignored vendor adapters but left git status clean — yet AGENTS.md's lightweight-mode note, read literally ("did a file change?"), seemed to demand a lite session log. It shouldn't: SKILLS.md already says the sync op "touches no committed file… not a version change," and the note's own anchor is the git diff. The wording just didn't say tracked.

Changed

1. Lightweight-mode test re-keyed to a tracked change. AGENTS.md (root + template) now states the objective test is the git diff over tracked files, not any filesystem write. The Read-only tier explicitly covers a run whose only writes are gitignored, regenerated artifacts — sync skill adapters, review-scratch/ snapshots, the compiled lint artifact — as no session log. The middle tier reads "A tracked file changed…"; the closing line, "anything that touched a tracked file." Aligns the note with SKILLS.md (sync touches no committed file) and prevents a spurious lite log after every adapter sync. Wording-only — no file shape, skill, or script change.

Lockstep: AGENTS.md (root + template), VERSION→4.10.3, UPGRADE.md 4.10.2→4.10.3 rung + table, README table. SKILLS.md / DECAY.md / REVIEW.md / .agent/schema.md unchanged.

---

Version 4.10.2, 6/18/2026

Fresh-context-review critique fixes (PATCH). The v4.10.x line was put through its own second-opinion ritual: a clean-vendor reviewer (GitHub Copilot CLI) challenged the milestone and returned a critique applied through the apply-critique loop. It confirmed the design was sound, cleared the no-build-step-agent-run question, and surfaced four scoped refinements — two documentation-consistency gaps that would have re-triggered the same review noise, one real parser robustness hole the v4.10.1 fix didn't cover, and one expectation-setting caveat.

Fixed

1. memory-lint — guard against an unclosed footer. FOOTER_RE is now bound to a single line ([^\n], no re.S). A malformed footer (<!-- id: foo | ... with no closing -->) can no longer let the non-greedy field capture span newlines and swallow the rest of the file up to a stray -->, which would have silently misparsed tier/superseded fields and corrupted decay counts with no error. Same lesson as v4.10.1: a verifier whose job is trustworthy counting must not be fooled by malformed input. Script-only; no description change → adapters unchanged.

Changed

1. Install protocol warns before overwriting a locally-modified built-in. ENABLE.md §5i (and the UPGRADE.md 4.10.0 / 4.10.2 rungs) now say: before overwriting an already-installed tool-managed built-in, diff it against the source; if it was locally customized, warn the human and let them decide rather than silently discarding their change. Makes the tool-managed-copies contract a checked step, not convention-only. Agent-run at the human's direction (no-build-step-agent-run); a no-op on a fresh enable.
2. upgrades-additive invariant text carries its exception inline. The invariant in this repo's memory/continuity.md now states the tool-managed-built-ins carve-out at its declaration (it was only in ENABLE/UPGRADE prose), so a reader encountering it cold no longer re-flags the overwrite as a contradiction — the exact noise the dogfood reviewer raised.
3. second-opinion — same-vendor vs. different-vendor caveat. A Notes bullet now states that a same-vendor clean session (or spawned subagent) tests the mechanism, while a different vendor adds the epistemic diversity that is the whole point of a fresh reviewer for high-stakes milestones. Body only — description unchanged → adapters unchanged.

Deferred (reviewer rated low-confidence/contrived): hardening pinned_open_threads against deeply nested lists — tracked as an Open Thread, not changed, since current session-log formatting is flat.

---

Version 4.10.1, 6/18/2026

memory-lint bug fix — line-anchor the Memory-References detection (PATCH). Found while running the verifier during a memory review: a first pass false-positived [over-archived] sync-adapters-v420. The review's own session log quoted the heading string "## Memory References" in its prose (describing the very check), and the script's un-anchored find() matched that inline mention before the real section — scooping the neighbouring ## What happened paragraph (which named the archived id) in as the references block. The archival was in fact correct. A verifier whose whole job (v4.9.0) is trustworthy decay-counting must not be fooled by prose, so this anchors the heading match to the start of a line.

Fixed

1. agent-skills/memory-lint/scripts/memory-lint.py — memref_ids() now anchors the heading to a real line ((?m)^## +Memory References[ \t]*$) and bounds the block at the next line-anchored heading, instead of text.find("## Memory References"). A session log may now safely quote the heading inline without producing a false over-archived error. Verified with added cases: the false-positive is gone, genuine references in the real section are still detected, and block-bounding still stops at the next heading. No description change → adapters unchanged; the tool still never runs it (no-build-step-agent-run).
2. Ignore Python bytecode caches — .gitignore and templates/.gitignore now ignore __pycache__/ + *.py[cod]. The bundled memory-lint helper generates these on run (in this repo and in every enabled target as of v4.10.0); the .py source under agent-skills/ stays tracked. Add-only, consistent with the v3.1.0 .gitignore propagation.

---

Version 4.10.0, 6/18/2026

Fresh-context second opinion — a clean-memory reviewer as a deliberate gate. A long session accumulates self-trust; the agent that built a solution over-trusts its own trajectory. This release adds a skill pair for the highest-value antidote: at a milestone or risk point, hand a compact snapshot to an AI with clean memory (a fresh session or a different vendor) and ask it to challenge the work. It folds an external brainstorming idea (the "Agent Interchange Format" draft) into the existing skills layer + VBDI rather than a standalone spec — the net-new surface is small: a security advisory on export, the handoff ritual, and the critique shape. Everything else reuses continuity.md + session logs, the VBDI human gate, and memory-lint/build-tests. It carries the hard-won v4.8/v4.9 lesson in its bones: the fresh reviewer is a hypothesis generator, not an authority (a clean-context reviewer once over-archived still-referenced facts here), so critique is advisory and gated by deterministic checks + a human.

Added

1. agent-skills/second-opinion/ — generates a review snapshot derived from continuity.md + recent session logs (decision-relevant deltas only — never a parallel committed state file), behind a security advisory the human must acknowledge before any state is exported. Supports milestone mode (challenge a done-looking milestone) and reactive mode (blocked / uncertain / risky). Writes to gitignored review-scratch/.
2. agent-skills/apply-critique/ — consumes the reviewer's critique through a bounded, validated, human-gated loop: parse → plan a few scoped fixes → apply → validate (build/tests + memory-lint if memory changed) → summarize applied vs. rejected. Conflicts become Open Threads (never-pick-a-winner); the cycle writes a normal session log.
3. review-scratch/ — a gitignored, per-machine scratch dir for snapshots/critiques, with a README marking the files personal. Sharing one is a conscious human decision. Added to the root and templates/.gitignore; documented in .agent/schema.md.

Changed

1. Built-in skills now installed (ENABLE.md Step 5i + the 4.10.0 upgrade rung). Every enabled repo gets the built-in skills — second-opinion + apply-critique and memory-lint — copied into agent-skills/ with adapters regenerated; review-scratch/ added to the target .gitignore. memory-lint is installed too because the review ritual relies on it — this supersedes its v4.9.0 "tool-local / not auto-installed" stance. Install is idempotent and never touches unrelated agent-skills/ content; a fresh Mode A enable now creates a populated (never empty) agent-skills/. Report + scope + verify notes updated.
2. VERSION → 4.10.0; UPGRADE.md 4.9.0→4.10.0 rung + table; README version table. AGENTS.md / SKILLS.md / DECAY.md / REVIEW.md unchanged — the critique→repair loop reuses the existing ritual. Design: docs/DESIGN-fresh-context-review.md.

---

Version 4.9.0, 6/18/2026

memory-lint — a deterministic verifier skill (move the decay arithmetic off the LLM). v4.8.0 added a markdown self-verify step, but the primary count was still hand-done — and an LLM (Copilot, then a hand re-check) miscounts. This release ships the stronger fix Copilot argued for: a portable, optional memory-lint skill whose bundled script deterministically checks memory integrity. The agent judges meaning; the script does the counting. On its very first run it caught a real over-archival that two rounds of human review had missed (skills-layer-v411-fixes, last referenced 16 sessions ago ≤ archive_window 20) — now reactivated.

Added

1. agent-skills/memory-lint/ — a portable skill (neutral SKILL.md + scripts/memory-lint.py, Python 3 stdlib, no install). Deterministic checks: no id in both continuity.md and the archive; no archived-as-faded fact referenced within archive_window (the decay-miscount guard); advisory overdue-for-archival (excluding core / superseded / pinned - [ ] threads); supersession links resolve. Exit non-zero on error → wire to a pre-commit hook / CI. Optional and agent/human-invoked (no-build-step-agent-run — the tool never runs it); it lints the arithmetic, not the meaning.

Changed

1. REVIEW.md step 6 ("Verify archival") now points to memory-lint as the recommended deterministic version of the check ("let the script count"). VERSION → 4.9.0; UPGRADE.md 4.8.0→4.9.0 rung + table; README. AGENTS.md / SKILLS.md / DECAY.md unchanged.

Fixed

1. Reactivated skills-layer-v411-fixes — memory-lint flagged it as over-archived by the 2026-06-18 Copilot review (sslu 16 ≤ archive_window 20); moved back to continuity.md (active).

---

Version 4.8.0, 6/18/2026

Review self-verify guard — catch decay miscounts before they archive. A GitHub Copilot CLI review (2026-06-18) over-archived three recent, active facts — it miscomputed sessions_since_last_used (counted ~5–9 as ">20") and ignored its own "still referenced in review window" note. The decay rule was right; what was missing was a check that the count was. Counting session files by hand is the easiest review step to get wrong, and archival is the costliest error — so the review now verifies its own archival against a cheap, grep-able signal.

Added

1. "Verify archival" step in REVIEW.md (new step 6, before stamping): for each fact about to be archived as faded, grep the last archive_window session files for its id — any hit means the count was wrong, so keep the fact (it's active/archive-candidate), don't archive. Then confirm no id lives in both continuity.md and the archive. Superseded facts are exempt (they archive on truth-state, not recency). Added an Archive-verify: line to the review-summary format. Pure markdown; replaces a hand-counted judgment with a checkable signal for the riskiest operation.

Changed

1. VERSION → 4.8.0; UPGRADE.md 4.7.1→4.8.0 rung (re-sync REVIEW.md) + table; README. AGENTS.md / SKILLS.md / DECAY.md unchanged.

---

Version 4.7.1, 6/17/2026

Lightweight mode — sharpen the skip criterion to an objective test. v4.7.0 let any "memory-neutral" task drop to a lite log. But "trivial" is a judgment call — both AI and human misjudge it, and a misjudged change that actually mattered would slip out of the ledger. So the distinction is now keyed to the objective question "did a file change?", not a subjective "is this trivial?": - Read-only (no file changes — orientation, Q&A, exploration) → no session log. - Any file change, however small (a one-line fix, a typo) with no memory-relevant event → a one-line "lite" session log (never skipped on a "felt trivial" call; the git diff anchors it). - Memory-relevant event → the full ritual.

Changed

1. AGENTS.md (root + template) "Lightweight mode" note rewritten to the three-tier, file-change-keyed form above. VERSION → 4.7.1; UPGRADE.md 4.7.0→4.7.1 rung + table; README. SKILLS.md / DECAY.md / REVIEW.md unchanged.

---

Version 4.7.0, 6/17/2026

Lightweight mode — scale the ceremony to the memory impact. From a real cross-vendor enablement: Kiro used the installed tool to AI-enable a new source-code repo and gave an honest assessment — the protocol "helps more than it interferes" (immediate orientation from memory/, actionable SKILLS.md, real multi-agent-continuity value), but the per-session write ceremony is disproportionate for trivial tasks ("for a two-line script and a skill file, the protocol's weight is noticeable") and "could benefit from a lightweight mode." This release adds exactly that — without breaking the event-sourced ledger.

Added

1. Lightweight mode (memory-neutral tasks) in AGENTS.md (root + template), in "After Every Session": if a task produced no memory-relevant event (no new/changed fact, no decision, no Open Thread touched, no project-state change), write a one-line "lite" session log (persist-time filename + **Agent:** + a lightweight-marked summary + ## Memory References → (none)) and skip the full template, fact-footer bookkeeping, and continuity edits. The ledger stays continuous (every session still logged), so multi-agent continuity is preserved; the review counts a lite log like any session but it carries no references, so usage is unaffected. When unsure, write the full log; Vision/Blueprint/invariant/supersession changes are never memory-neutral. DECAY.md / REVIEW.md need no change — a lite log is just a session file with no ## Memory References.

Changed

1. VERSION → 4.7.0; UPGRADE.md 4.6.0→4.7.0 rung (re-sync AGENTS.md) + table; README. SKILLS.md / DECAY.md / REVIEW.md unchanged.

---

Version 4.6.0, 6/17/2026

Vendor-neutral commit attribution — encode the convention once, every vendor follows. A Kiro observation crystallized it: Claude Code adds a deliberate, human-gated Co-Authored-By: trailer automatically (it's in its harness), Kiro needed a per-machine hook + steering to do the same, and Gemini/Cursor do nothing by default. That's exactly the gap the shared layer closes — so AGENTS.md now extends its existing "identify yourself" principle (already true for session logs) to commits: any vendor's agent gets Claude's behavior with no per-vendor hook. Serves multi-contributor traceability (bp-multi-user) + provenance.

Added

1. Commit-attribution convention in AGENTS.md (root + template), in the "After Every Session" step 4 + checklist: "Commits are deliberate and human-initiated. When you commit at the human's direction, identify yourself the same way you do in session logs — e.g. a Co-Authored-By: <your agent name> trailer — so authorship is traceable across vendors. (If your runtime already adds one, nothing to do.)" Soft by design — guides, doesn't prescribe git workflow; a no-op for runtimes (like Claude Code) that already do it.

Changed

1. VERSION → 4.6.0; UPGRADE.md 4.5.2→4.6.0 rung (re-sync AGENTS.md) + table; README. SKILLS.md / DECAY.md / REVIEW.md unchanged.

---

Version 4.5.2, 6/17/2026

Kiro hooks in Mode C + a bootstrap edge-case note — from a Windows/Kiro enable. Cloning an agent-memory repo on a fresh machine and opening it in an enterprise Kiro surfaced two things: (1) .kiro/ doesn't exist on a fresh clone (gitignored, per-machine), and an enterprise IDE may self-bootstrap from its own onboarding/MCP before reading AGENTS.md — the fix is the human nudge "Start from AGENTS.md"; and (2) once Kiro is running it deposits hooks and steering into .kiro/ (e.g. a commit-signature hook). Those stay gitignored/per-machine and don't touch the shared memory/ layer, but the migration protocol didn't cover hooks for other repos being enabled. (The Kiro commit hook is human-gated — it fires only when the human says "commit" and adds a co-author trailer — so it aligns with agent-memory's deliberate-commit model; no tension.)

Changed

1. MIGRATE.md — the Kiro per-vendor protocol gains a Hooks sub-case: .kiro/hooks/*.kiro.hook are preserved verbatim under legacy/kiro/hooks/, never converted or run (automation is a human, gated decision). Human-gated commit hooks (like Kiro's, triggered by the human saying "commit") align with agent-memory and need no action; only a hook that commits or pushes unprompted is surfaced as an Open Thread (never-pick-a-winner).
2. README.md — a bootstrap edge-case note: on a fresh clone the vendor dirs are gitignored, so an enterprise IDE may self-bootstrap first; tell the agent "Start from AGENTS.md", then run sync skill adapters. Enterprise hooks/steering in .kiro/ stay gitignored and per-machine.
3. VERSION → 4.5.2; UPGRADE.md 4.5.1→4.5.2 rung + table. AGENTS.md / SKILLS.md / DECAY.md / REVIEW.md unchanged.

---

Version 4.5.1, 6/17/2026

Skills-layer guidance — from a cross-machine Gemini CLI dogfood. A real Gemini CLI run (different machine) surfaced two rough edges, both wording/guidance, no shape change: (1) after sync, a natural-language "run hello-world" kept reading agent-skills/ instead of the freshly-generated .gemini/commands/ command; (2) the agent told the user to commit the vendor adapter dirs. Root causes: Gemini custom commands are slash commands (/<name>, explicit) — not natural-language auto-triggers like Claude/Cursor/Kiro — so NL correctly routes through the AGENTS.md baseline to the same neutral skill (identical result, not a defect); and our docs stated adapters are gitignored but never explicitly said don't commit / recommend committing them.

Changed

1. SKILLS.md — the Gemini adapter is now documented as a slash command /<name> (explicit, not NL-matched); added a "trigger semantics differ per vendor" note (Claude / Cursor / Kiro are description-matched, Gemini is slash-invoked, all pointing to the same neutral skill); and a never-commit-the-adapters guard on the sync skill adapters operation (report adapters as "gitignored — do not commit; only agent-skills/ is shared").
2. AGENTS.md (root + template) — the adapter line now says "never commit them".
3. docs/DESIGN-skills-layer.md — recipe table flags Gemini as slash-invoked; a trigger- semantics + commit-guard note under the table; status line. VERSION → 4.5.1; UPGRADE.md 4.5.0→4.5.1 rung + table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.5.0, 6/16/2026

Kiro support — convergence on open standards. Amazon's Kiro IDE adopts the two open standards this tool already bets on: it auto-reads a root AGENTS.md (so the memory layer works with no pointer file) and its Agent Skills follow the open Agent Skills standard (the same SKILL.md shape as Claude). So enabling a Kiro repo needs almost nothing new — this release adds a 4th skills adapter target and teaches Mode C to detect/migrate Kiro's repo-local artifacts. Kiro Powers need no special handling: they are partner-published bundles (MCP + steering + hooks) that consume open-standard skills, which ours already are — the tool never emits a Power.

Added

1. Kiro skills adapter (SKILLS.md recipe): a 4th target .kiro/skills/<name>/SKILL.md — identical shape to the Claude adapter (frontmatter name + description + a pointer body), since Kiro follows the open Agent Skills standard. sync skill adapters now writes four adapters; the standing read-only Skills adapter check (UPGRADE.md) now scans for it too.
2. Kiro in the Mode C detection table (MIGRATE.md) + a per-vendor Kiro protocol: .kiro/steering/*.md → memory/instructions.md; .kiro/skills/ → promoted to agent-skills/ (open-standard, same as Claude); .kiro/specs/ preserved verbatim under legacy/kiro/specs/ (never auto-flattened — folding a spec into the Vision/Blueprint is a human, altitude-gated decision).

Changed

1. Adapter lists across the living docs now include .kiro/skills/ (AGENTS.md root + template, ENABLE.md Step 5h / Step 8 / scope, docs/DESIGN-skills-layer.md recipe table + Option A). .gitignore needs no change — .kiro/ was already in the managed block; the comment now names .kiro/skills/ among the adapters. VERSION → 4.5.0; UPGRADE.md 4.4.0→4.5.0 rung + table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.4.0, 6/16/2026

Lightweight skills — conscious, not per-session. Skill creation is a deliberate, occasional developer action, so it leaves the per-session path. The per-session AGENTS.md now carries only the runtime baseline (read & follow a matching skill) + a pointer; the adapter recipe and the sync / adopt / sanity-check operations move to a new on-demand SKILLS.md (installed like DECAY.md/REVIEW.md). The v4.3.0 per-session "skills safety check" is removed; upgrades instead do a read-only filename check that recommends sync skill adapters if an adapter is missing/orphaned. Trims the per-session bootstrap (the skills recipe was ~1.3K tokens read every session).

Added

1. SKILLS.md (installed at repo root) — the on-demand skills reference: authoring convention, adapter recipe, and the sync / adopt / skill sanity check operations.
2. Standing "Skills adapter check" in UPGRADE.md — read-only, filename-only; runs on any Mode B re-enable (incl. "up to date") and recommends sync if adapters are missing/orphaned.

Removed

1. The per-session "Skills safety check" step from "After Every Session" (both AGENTS.md files) + its checklist line — superseded by the on-demand + upgrade-time model.

Changed

1. AGENTS.md (root + template): "Skills" reduced to the runtime baseline + a pointer to SKILLS.md. ENABLE.md (Step 5h references SKILLS.md; Step 6 installs it; Step 8 verifies), MIGRATE.md, .agent/schema.md now point at SKILLS.md. VERSION → 4.4.0; UPGRADE.md 4.3.3→4.4.0 rung + standing check + table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.3.3, 6/16/2026

Skill description guidance (reviewed from external feedback). A description is a model-matched activation signal read within a small discovery budget, so it should be a concise, trigger-phrase-rich summary — not a long abstract paragraph (which weakens activation). Also clarified that YAML >/| folded/literal blocks are YAML-only and don't carry into the TOML adapter, so the canonical value stays one logical line. Wording only.

Changed

1. AGENTS.md "Authoring a skill" asks for a concise (~1–2 sentence), trigger-rich description; the recipe notes the >/| YAML-only caveat.
2. Tightened the dogfood hello-world description (~35 words) and regenerated its adapters.

VERSION → 4.3.3; UPGRADE.md 4.3.2→4.3.3 rung + table; README; docs/DESIGN-skills-layer.md §9 note. DECAY.md / REVIEW.md unchanged.

---

Version 4.3.2, 6/16/2026

Skill description hardening, surfaced by a deliberate skill-lifecycle sanity check (dogfood). The check confirmed the mechanics (git split / pointer integrity / per-vendor format / adopt / prune all correct) but found two hard-to-spot description hazards. Wording/clarity only — no shape, data, or behavior change.

Changed

1. Adapter description mirrors the neutral skill's verbatim — never abbreviate it (the prior hand-made hello-world adapters had abbreviated descriptions that silently drifted from the skill).
2. Skill descriptions must be single-line and quote-free — a description containing a " would otherwise emit invalid TOML (Gemini) or .mdc/YAML frontmatter (Cursor) when synced. The recipe adds an escape/quote fallback for unavoidable cases.
3. Fixed the dogfooded hello-world skill's description (dropped inner quotes) and regenerated its adapters to mirror it.

VERSION → 4.3.2; UPGRADE.md 4.3.1→4.3.2 rung + table; README. templates/AGENTS.md (+ root AGENTS.md) re-synced; docs/DESIGN-skills-layer.md §9 note; DECAY.md / REVIEW.md unchanged.

---

Version 4.3.1, 6/16/2026

Skills-layer doc fixes, surfaced by a fresh-agent test-drive of the v4.3.0 session-close ritual (dogfooded on a real target). Wording/clarity only — no shape, data, or behavior change.

Changed

1. "Adopt a skill" no longer instructs a mid-ritual commit — it conflicted with the session-close ritual (which only reminds to commit). Adopt now says to stage the neutral skill for the session-end commit; the agent doesn't self-commit mid-ritual.
2. Session-close skills check notes that adoption (which changes the tree) should run before writing the session log, so the log records it.
3. Clarified adopt body normalization (preserve the procedure, neutralize vendor phrasing) and the detection locations.

VERSION → 4.3.1; UPGRADE.md 4.3.0→4.3.1 rung + table; README. templates/AGENTS.md re-synced; DECAY.md / REVIEW.md unchanged.

---

Version 4.3.0, 6/16/2026

Skill authoring convention + "adopt skill" safety-net. Closes a footgun in the skills layer: the source of truth is agent-skills/<name>/SKILL.md, but a user — or a vendor's built-in skill creator — might author a skill directly in a vendor folder (.claude/skills/, …), which is gitignored, so it's never shared and isn't the source of truth. v4.3.0 makes the authoring convention explicit and adds a reverse "adopt" operation, checked at session close. Additive (docs/protocol only).

Added

1. Authoring convention (AGENTS.md "Skills" → "Authoring a skill"): create skills in agent-skills/<name>/SKILL.md, then "sync skill adapters"; never author in a vendor folder.
2. "Adopt a skill" operation (AGENTS.md "Skills" → "Adopt a skill"): promote a skill authored natively in a vendor folder into agent-skills/ (the reverse of sync — the same move migration makes at enable), then sync. The on-demand complement to migration's one-time promote.
3. Session-close safety check: the "After Every Session" ritual gains a step (+ checklist line) that detects a skill stranded in a vendor folder and prompts adoption before commit.

Removed

N/A.

Changed

1. AGENTS.md (root + template), MIGRATE.md (on-demand-adopt note), .agent/schema.md, docs/DESIGN-skills-layer.md. VERSION → 4.3.0; UPGRADE.md 4.2.0→4.3.0 rung + table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.2.0, 6/16/2026

"Sync skill adapters." Closes a gap surfaced by a real cross-machine test: the neutral agent-skills/<name>/SKILL.md travels via git, but the per-vendor adapters are gitignored and don't travel — so a freshly cloned/pulled repo has no native /-command / auto-trigger until they're regenerated locally, and re-running "AI enable" on an up-to-date repo is a no-op. v4.2.0 adds an on-demand operation to regenerate them, and moves the adapter recipe into the installed AGENTS.md so a target's own agent (any vendor) can self-sync. Additive (the runtime baseline always worked without adapters).

Added

1. "Sync skill adapters" operation — for each agent-skills/<name>/SKILL.md, regenerate the Claude/Gemini/Cursor adapters (idempotent) and prune orphaned ones. Invoked by saying "sync skill adapters"; touches no committed file (adapters are gitignored).
2. Canonical adapter recipe relocated to the installed AGENTS.md "Skills" section, so a contributor's agent — on any vendor, on any machine — can self-sync by reading its own AGENTS.md. ENABLE.md Step 5h now references this single recipe (DRY).

Removed

N/A.

Changed

1. AGENTS.md (root + template): "Skills" section gains the adapter recipe + the sync operation. ENABLE.md Step 5h references the recipe instead of duplicating it. .agent/schema.md notes on-demand sync. VERSION → 4.2.0; UPGRADE.md 4.1.1→4.2.0 rung
2. version table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.1.1, 6/16/2026

Skills-layer refinements — pre-adoption corrections to v4.1.0 (which shipped the same day and was not yet consumed by any repo), so the first real target run starts clean. The adapter formats were verified against current vendor docs in the process.

Changed

1. Folder renamed skills/ → agent-skills/. skills/ is a common top-level dir; agent-skills/ is namespaced to avoid collision (and stays visible/discoverable, a peer to memory/). All living docs updated; vendor .claude/skills/ references unchanged.
2. Cursor adapter fixed. .cursor/rules/<name>.mdc now emits the agent-requested rule type — description + empty globs: + alwaysApply: false — verified against current Cursor docs. (Gemini .gemini/commands/*.toml with description/prompt/ {{args}}, and Claude .claude/skills/<name>/SKILL.md, were verified correct as shipped.)
3. Collision guard. Enable/migrate never overwrite a pre-existing agent-skills/ — they raise a Contradiction: Open Thread (never-pick-a-winner) and stop.
4. Vendor-dir double-duty clarified (MIGRATE.md): .cursor/rules/ and .gemini/ are both migration sources and adapter targets — archive originals to legacy/ first, then generate adapters, so they never collide.

VERSION → 4.1.1; UPGRADE.md 4.1.0→4.1.1 rung + version table; README. DECAY.md / REVIEW.md unchanged.

---

Version 4.1.0, 6/15/2026

Cross-vendor skills layer. The shared layer gains its third leg — capabilities — beside memory and steering. A skill is committed, vendor-neutral markdown (skills/<name>/SKILL.md: a name, a description = when-to-use, a procedure, optional scripts). The AGENTS.md "Skills" section is the universal runtime (the agent reads the skill — works on any vendor); native adapters (.claude/skills/, .gemini/commands/, .cursor/rules/) are thin, regenerated, gitignored pointers. Migration promotes vendor skill bundles (e.g. .claude/skills/) into skills/ rather than flattening them into steering. Additive (a repo with no skills is unchanged). Design: docs/DESIGN-skills-layer.md.

Added

1. skills/ layer — neutral, committed skills/<name>/SKILL.md capabilities (the shared source of truth), documented in .agent/schema.md.
2. AGENTS.md "Skills" baseline (root + template) — the agent-as-runtime mechanism; works on any vendor with no per-vendor engine.
3. Per-vendor adapters — generated Claude (.claude/skills/), Gemini (.gemini/commands/), and Cursor (.cursor/rules/) pointers; regenerated, gitignored.
4. Migration promotion — MIGRATE.md Section B2 + the Claude Code skills protocol: detect .claude/skills/, promote into skills/ (originals preserved under legacy/), regenerate adapters. ENABLE.md Step 5h defines adapter generation.

Removed

N/A.

Changed

1. ENABLE.md (Step 5h + verify/report/scope), MIGRATE.md (principle 6, Section B2, Claude protocol, detection table, continuity note), AGENTS.md (root + template), .agent/schema.md, templates/.gitignore (comment), README. VERSION → 4.1.0; UPGRADE.md 4.0.0→4.1.0 rung + version table. DECAY.md / REVIEW.md unchanged.

---

Version 4.0.0, 6/15/2026

The forward layer. agent-memory gains a cognitive lifecycle loop on top of its memory substrate: Current State → Vision → Blueprint → Design → Implementation → Feedback. Memory was backward-looking (faithful to what happened); VBDI makes it goal-aware (faithful to what was intended) — the "predictable innovation with human partnership" mission. Integrates an independently-drafted cognitive framework (docs/agent-cognitive-framework.md), designed in docs/DESIGN-vbdi-lifecycle.md. Additive (a repo with no Vision is unchanged) — bumped to 4.0.0 to mark the new layer, the forward counterpart to 3.0.0's evolving-memory layer.

Added

1. memory/vision.md — the north-star artifact (templates/memory/vision.md): target state, for whom, success criteria, non-goals. core, invariant-verified, one per repo.
2. Blueprint — the Vision↔Current-State gap as typed (blueprint) Open Threads (… → serves: <vision-id>); no new file.
3. Altitude trace — Implementation → Design → Blueprint → Vision, linked by id; a broken link is drift (DECAY.md §10/§12), grep-detectable.
4. Bootstrap, never fabricate — ENABLE.md (Step 5g) and the UPGRADE.md 3.7.0→4.0.0 rung create a ⚠️ DRAFT Vision (Current-state context only; target left for the human) and raise a (vision-bootstrap) human gate.

Removed

N/A.

Changed

1. DECAY.md (§10 altitude drift + new §12 "The forward layer"), REVIEW.md (Vision in invariant-verification + altitude drift in the backstop), .agent/schema.md, AGENTS.md (root + template: a "cognitive loop" section + Vision in the session read-list), ENABLE.md (Step 5g + verify + report). VERSION → 4.0.0; UPGRADE.md rung + version tables; README. The tool dogfoods its own confirmed Vision + Blueprint.

---

Version 3.7.0, 6/14/2026

Provenance + retrieval-at-scale — the last backlog item. Event-sourcing already records which session each fact came from; v3.7.0 surfaces it as an optional origin: footer, and documents retrieval as deliberately lexical + indexed (grep + archive/INDEX.md + origin pointers), with vector/semantic retrieval intentionally out of scope. Closes the assessment's provenance gap (#6); addresses the retrieval gap (#4) by design.

Added

1. Optional origin: <session-file> footer field (DECAY.md §1 + new §11) — the session where a fact was Created; set at creation, repairable by review (the earliest Created event). Provenance becomes one hop, and is the cheap defence against memory poisoning (every fact traces to an immutable session).
2. DECAY.md §11 "Provenance & retrieval" documents the lexical+indexed retrieval strategy (grep continuity → grep archive/INDEX.md → follow origin → optional sessions/INDEX.md), bounded by project scale.

Removed

N/A.

Changed

1. AGENTS.md (root + template) set origin on new facts and point at the retrieval path; .agent/schema.md documents origin + the retrieval note; REVIEW.md notes origin backfill; ENABLE.md notes origin is omitted at enable. The worked example shows origin on the facts created in its session.
2. VERSION → 3.7.0; UPGRADE.md 3.6.0→3.7.0 rung + version tables; README. Assessment gap #6 flipped ⬜ → ✅; gap #4 (lexical retrieval) marked ◐ by-design; Retrieval scorecard row updated.

---

Version 3.6.0, 6/14/2026

Memory smoke test — a cheap, no-code answer to "memory evaluation is unsolved": a memory/smoke-test.md whose questions a fresh agent should be able to answer from the memory layer alone. A ❌ is a memory gap to fix, not a question to soften. Closes the assessment's gap #5. Additive (a new installed file).

Added

1. templates/memory/smoke-test.md — generic orientation questions (project & type, stack, invariants, recent decisions + why, open threads, conventions, preferences, supersessions) plus project-specific questions seeded at enable, a how-to-run, and a result log.
2. ENABLE.md step 5f generates it (seeding 2–4 project-specific questions from the analysis); Step 8 verifies it; Step 9 reports it. REVIEW.md notes that a review is a natural time to run it.

Removed

N/A.

Changed

1. .agent/schema.md documents the file. VERSION → 3.6.0; UPGRADE.md 3.5.0→3.6.0 rung + version tables; README. Assessment gap #5 flipped ⬜ → ✅ and the Evaluation scorecard row ⬜ → ✅.

---

Version 3.5.0, 6/14/2026

Write-time contradiction check — the migration-time "surface contradictions, never pick a winner" rule now applies to normal sessions: a new fact is scanned against existing ones before it's recorded. Closes the assessment's gap #3 (the last "Partial"). Additive — a behavioral rule + a review backstop; no new fields or knobs.

Added

1. DECAY.md §10 — write-time contradiction check: on adding/rewriting a fact, scan core/invariants + active decisions in the same area; a clear replacement → supersede (§9), a genuine conflict → a Contradiction: Open Thread, a clash with a core invariant → stop and surface.
2. A contradiction backstop in REVIEW.md (the review scans the facts it already reads and flags conflicts it finds).
3. The before-adding-a-fact check in the after-session step of AGENTS.md (root + template).

Removed

N/A.

Changed

1. VERSION → 3.5.0; UPGRADE.md 3.4.0→3.5.0 rung + version tables; README. Assessment gap #3 flipped ⬜ → ✅ and the "Truth maintenance" scorecard row ◐ → ✅.

---

Version 3.4.0, 6/14/2026

Invariant verification — never-decay facts (core / Architectural Invariants) can quietly go "confidently wrong" when circumstances change. The review now periodically prompts a human to re-confirm them (or supersede the false ones via v3.3.0). Closes the assessment's gap #2. Additive (a policy knob + a tracker field); already-enabled repos catch up via the UPGRADE.md 3.3.0→3.4.0 rung.

Added

1. verify_invariants_every (default 20) in decay-policy.md, and a last_invariant_check tracker in continuity.md Project State.
2. Review routine step 6 (REVIEW.md): when due, raise one Open Thread asking a human to re-confirm every never-decay fact, then stamp last_invariant_check. The review never auto-invalidates — it only prompts; the human confirms or supersedes (§9). An Invariants: line joins the review summary.
3. The worked example shows a first invariant-verification prompt.

Removed

N/A.

Changed

1. DECAY.md §6 gains a "never-decay ≠ never-checked" note; .agent/schema.md documents the last_invariant_check field + the knob. VERSION → 3.4.0; UPGRADE.md 3.3.0→3.4.0 rung + version tables; README. Assessment gap #2 flipped ⬜ → ✅.

---

Version 3.3.0, 6/13/2026

Supersession / fact-invalidation — the evolving-memory layer can now represent a fact becoming false (a decision reversed, a dependency dropped), not just fading from disuse. Closes the industry-alignment assessment's one "High" gap. Additive (optional footer fields + a terminal tier); already-enabled repos pick it up via the UPGRADE.md 3.2.0→3.3.0 rung.

Added

1. A terminal superseded tier and optional superseded-by / supersedes footer fields (DECAY.md §9). When a fact is reversed/invalidated, the agent marks it superseded immediately, adds the successor (supersedes: <old>), and records Superseded: <old> → <new> in the session log's ## Memory References.
2. Review behavior: superseded facts are archived flagged "superseded" (distinct from "faded"), promptly (no archive_window wait), links preserved in the archive + INDEX.md; a Superseded: N line joins the review summary. Superseded is terminal — never reactivated by a reference (only a human can reverse it).
3. A worked supersession in examples/evolving-memory-example/ — a reversed REST-versioning decision threaded through before/after/session-log/archive.

Removed

N/A.

Changed

1. DECAY.md, REVIEW.md, .agent/schema.md, and AGENTS.md (root + template) updated for the above; VERSION → 3.3.0; UPGRADE.md 3.2.0→3.3.0 rung + version tables; the industry-alignment assessment flips gap #1 ⬜ → ✅.

---

Version 3.2.0, 6/13/2026

Protocol clarifications from a real-work field report (a Node.js→Rust refactor of an enabled repo, driven entirely by that repo's own memory layer). Documentation and wording only — no memory-file shape change; already-enabled repos pick these up via the UPGRADE.md 3.1.0→3.2.0 rung.

Added

1. After-session checklist in AGENTS.md (root + template), plus an optional Stop/pre-commit hook recipe (docs/optional-ritual-hook.md) to prompt the ritual without breaking the no-code core.
2. Stack-fact altitude rule: continuity.md → ## Stack & Tools is the canonical live home for language/deps/versions; instructions.md keeps a high-level descriptor and points there (notes added to both templates and .agent/schema.md).

Removed

N/A.

Changed

1. Session model defined precisely: a session is one log-write (a long conversation may produce several), and the title's start is now best-effort (the persist-time end is the required ordering stamp) — no more fabricated starts.
2. Metadata ownership pinned and a contradiction fixed: ENABLE.md Step 5b seeded new facts as tier: active while DECAY.md/AGENTS.md said tier: working — now uniformly born working. The agent seeds id/created/tier/uses: 1; the review owns uses/last_used/tier thereafter.
3. Open-Thread archival rule surfaced to the agent: mark [x] and leave it for the review to sweep — don't self-archive.

---

Version 3.1.0, 6/13/2026

Propagate the AI-infrastructure .gitignore into enabled repos. Personal AI-IDE runtime directories (.claude/, .kiro/, .cursor/, …) are per-machine state that should never be committed to a shared repo; the tool's own repo already ignored them, but enablement never passed that design on to targets. Additive and backward-compatible; already-enabled repos catch up in place via UPGRADE.md Mode B.

Added

1. templates/.gitignore — the canonical managed block (sentinel-headed) listing the AI-IDE runtime directories and Aider local-history/cache files to ignore, while explicitly keeping steering files and the memory/ layer tracked.
2. UPGRADE.md rung 3.0.0 → 3.1.0 — creates or appends the .gitignore block in already-enabled repos (idempotent, add-only).

Removed

N/A — additive.

Changed

1. ENABLE.md Step 7 — now installs a .gitignore (copied from the template) when the target has none, and appends only the missing managed entries when one exists; never removes or reorders the user's existing entries. Previously it only added a comment to an existing file and refused to create one.
2. ENABLE.md — Step 8 verifies the .gitignore sentinel + entries; the Step 9 report and the Notes file-scope list now include .gitignore.
3. VERSION → 3.1.0; version tables in UPGRADE.md and README.md updated.

---

Version 3.0.0, 6/13/2026

Evolving memory layer. Facts stop being static: they carry usage metadata, fade through tiers, and archive when stale — recomputed deterministically from the session-log event stream. Breaking memory-file shape change (an un-upgraded agent couldn't correctly maintain the new files); already-enabled repos catch up in place via UPGRADE.md Mode B. Upgrades are additive and non-destructive.

Added

1. Evolving-memory layer — facts carry usage metadata (id, created, last_used, uses, tier) and fade through tiers (active → recent → archive), recomputed deterministically from session logs. Full design in docs/DESIGN-evolving-memory.md.
2. DECAY.md — metadata fields, tier lifecycle, and the deterministic integer decay rules (no floating-point scoring).
3. REVIEW.md — the review ritual that recomputes usage from session logs and archives faded facts.
4. ## Architectural Invariants section in memory/continuity.md — core-tier facts (the tool's safety philosophy) that never decay.
5. ## Memory References section in the session-log schema — the event source the review tallies usage from.
6. templates/memory/decay-policy.md — tunable decay windows (defaults 3/8/20, review every 10 sessions).
7. memory/archive/ — cold storage for faded facts (never deleted).
8. Versioning: root VERSION file (semver) plus a per-repo .agent/version.md install manifest recording what each enabled repo is on.
9. UPGRADE.md — the in-place version-upgrade ladder (reached only via ENABLE.md Mode B); tool-operator-only, not installed into targets.
10. examples/evolving-memory-example/ — the review ritual in action (continuity before/after, archive, a session log with Memory References).
11. Completed the root vendor bootstraps (.cursorrules, .windsurfrules, .github/copilot-instructions.md) and added .gitignore.

Removed

N/A — the upgrade is additive and non-destructive.

Changed

1. ENABLE.md — Mode B is now version-aware: an up-to-date repo is skipped, an older one is upgraded in place via UPGRADE.md; added the generate / install / verify steps for the evolving-memory files.
2. AGENTS.md (root + template) — the Before / During / After flow now references fact metadata and the review ritual.
3. .agent/schema.md — documents the metadata fields, the Memory References section, and the new files.
4. README.md and CLAUDE.md — added the architecture map, version table, and the evolving-memory section.

---

Version 2.0.0, 6/12/2026

Vendor detection and migration. The tool can now be pointed at a repo that already uses vendor AI files and fold them into the unified format, plus recognize and skip its own prior installs (idempotent re-runs).

Added

1. MIGRATE.md — a per-vendor detection table and conversion rules for 11 sources: Claude Code, Cursor, Cline, Roo Code, Aider, Continue.dev, Codeium/Windsurf, GitHub Copilot, GPT/Codex agents, Zed AI, and Gemini CLI.
2. Mode C (Migrate Vendor) — originals preserved under legacy/<vendor>/, steering folded into memory/instructions.md as ## Migrated rules from <vendor>, chat history converted to dated memory/sessions/ logs, and contradictions surfaced as Open Threads (the tool never picks a winner).
3. Mode B (Already Ours) — idempotent re-runs: an existing install is detected and skipped.
4. Dry-run support, so users can preview changes before committing.
5. Monorepo / multi-module guidance — detection signals, single-root default, and a Module Inventory template.
6. examples/rust-event-bus/ — a real Mode A fixture (the unedited output of enabling an actual Rust repo), replacing the earlier mock example.
7. examples/migrated-cursor-aider-project/ — a Mode C example with originals under legacy/ and converted session logs.
8. Documented the two memory layers (the repo's shared memory/ vs. the runtime's personal ~/.claude/).

Removed

1. The mock examples/node-project (taskflow-api) — replaced by the real rust-event-bus fixture.
2. The unreliable Duration field from the session-log schema.

Changed

1. Session-log filenames moved to YYYY-MM-DD-HHMMSS.md (UTC persist time); the title line is # Session (startZ - endZ) with full ISO 8601 millisecond timestamps — multi-contributor safe, with lexicographic sort matching chronological sort.
2. ENABLE.md — added version-drift guidance (prefer the build manifest as source of truth, surface drift as an Open Thread) and a post-report verification step.
3. MIGRATE.md — git-aware archiving (git mv for tracked files, plain move otherwise) and an integration quality gate that distinguishes project instructions (integrated into schema sections) from AI-only rules (appended verbatim).
4. Fixed a legacy/ .gitignore contradiction — legacy/ now commits to git, consistent with the "preserved" promise in README.md / MIGRATE.md.
5. Cold-start and accuracy refinements: inline project header in the CLAUDE.md / GEMINI.md templates, a User Preferences "explicit only, never infer" guard, and a sharpened different-agent check that compares agent family.

---

Version 1.0.0, 6/12/2026

Initial release. A no-code, markdown-only memory system that any AI agent can read, plus a Mode A enablement protocol to generate a tailored memory system in any repo.

Added

1. ENABLE.md — the enablement protocol (detect footprint → choose mode → analyze the repo → generate tailored memory files).
2. Mode A (Fresh Enable) — generate a tailored memory/ system from repo analysis.
3. templates/ — exactly what gets installed into a target repo: bootstrap files plus memory/ files with {{UPPER_SNAKE_CASE}} placeholders, and .agent/schema.md (the canonical memory-file format).
4. AGENTS.md — the dual-mode dispatcher: the memory protocol when working within the repo, the enablement protocol when AI-enabling another repo.
5. Per-vendor root bootstrap pointers — CLAUDE.md, GEMINI.md, .cursorrules, .windsurfrules, .github/copilot-instructions.md — each a thin pointer so any agent lands in the same AGENTS.md protocol.
6. This repo's own memory/ layer (instructions.md, continuity.md, sessions/) — the tool eats its own dog food.
7. The core safety constraints: target-repo scope only; never delete or overwrite vendor files; never modify source code or package manifests.

Removed

N/A — initial release.

Changed

N/A — initial release.