agent-memory · v4.26.1

A lightweight, vendor-neutral
AI memory + cognitive loop

Turning shared project context into deterministic delivery with near-zero drift — for human–AI collaboration, across any AI vendor.

Peer / leadership review · June 2026
The problem

Two failure modes of memory-driven agents

Memory drift & vendor silos

Context is re-explained every session; decisions are silently forgotten or contradicted; each AI tool hoards its own steering files — no shared, durable project memory.

Intent ↔ delivery gap

Agents are creative, but keeping what's built faithful to what was intended is unmanaged. Reversed decisions and "confidently wrong" facts pile up.

The bottleneck isn't the model — it's the architecture around it.

The idea

Two complementary layers over one shared memory/

Backward · Evolving Memory

"Where are we, and why?"

An event-sourced ledger projected into live state — deterministic decay, supersession, provenance. Faithful to what happened.

Forward · VBDI loop

"Where are we going — faithfully?"

A cognitive loop with an enforceable intent trace and human gates. Faithful to what was intended.

No-code, markdown-only. The files are the product; the agent is the runtime.

Backward layer

Evolving memory — event-sourced, deterministic

WRITEsessions (ledger) MANAGEreview ritual READcontinuity.md
Backward layer · trust

Memory you can trust

Supersession

A reversed/false fact is marked superseded — change, not just disuse.

Invariant re-verify

Never-decay facts are periodically re-confirmed by a human. Never-decay ≠ never-checked.

Contradiction check

A new fact is scanned against existing ones at write time — never pick a winner.

Provenance

Each fact traces to its source session — one hop; anti-poisoning by design.

Retrieval stays lexical + indexed by design — auditable and replayable, not a black-box vector store.

Forward layer

The VBDI cognitive loop

STATECurrent State TARGETVision GAPBlueprint SHAPEDesign ACTImplementation LEARNFeedback
The integration insight

Only two primitives are new

The loop mostly already exists in the memory layer — we integrate by mapping, not rebuilding.

PrimitiveRealized by
Current Statecontinuity.md
Vision newmemory/vision.md — the north star
Blueprint newtyped (blueprint) Open Threads = the gaps
DesignKey Decisions + Architectural Invariants
Implementationcode / commits, traced in sessions
Feedbackthe review ritual + decay + supersession
Why it's predictable

The trace is the determinism

Implementationserves → Designserves → Blueprintserves → Vision
v4.1.0 · capability layer

A third shared leg: cross-vendor skills

Memory and steering were already shared across vendors. Skills — reusable capabilities — are the third leg: one neutral, committed source of truth, runnable by any agent.

Neutral source

Committed agent-skills/<name>/SKILL.md — a name, a when-to-use description, a procedure. The single source of truth; travels with the repo.

Universal runtime

The AGENTS.md "Skills" baseline: when a task matches, the agent reads & follows the skill. The agent is the runtime — any vendor, no engine.

Thin adapters

Regenerated, gitignored pointers across six vendors — Claude, Gemini, Cursor, Kiro, GitHub Copilot & Google Antigravity — give native auto-trigger; never copies, so the neutral skill never drifts.

Migration promotes a vendor's skills (e.g. .claude/skills/) into the shared layer — originals preserved under legacy/, never flattened into steering.

capability layer · matured

Seven built-ins, one boundary

Installed into every enabled repo, tool-managed. They sort along the principle the layer learned the hard way: mechanize the arithmetic, leave the judgment to the agent.

Arithmetic · mechanized
  • memory-lint — 9-check integrity verifier (Python and Node)
  • refresh-metadata — recompute every fact's tier/usage
  • archive-fact — the archive move, truncation-proof
  • sync-adapters — regenerate the 6 vendor adapters
Judgment · stays with the agent
  • harvest-knowledge — distil durable facts from docs
  • second-opinion — snapshot for a clean-memory reviewer
  • apply-critique — bounded, human-gated apply loop
  • what to archive · how to resolve a clash · the Vision

Why: capable agents silently do half a multi-step ritual. The fix is structural — a runnable helper for the arithmetic + a lint advisory for the gap — never automating the decision.

fresh-context review

A second opinion from clean memory

A long session over-trusts its own trajectory. The highest-value antidote is a reviewer with clean memory — a fresh session or a different vendor that didn't live the work. Shipped as two built-in skills, installed into every enabled repo.

SNAPSHOTsecond-opinion security advisory clean-context reviewer APPLYapply-critique validate + human gate

Derived, not duplicated

The snapshot is distilled from continuity.md + recent sessions — never a parallel state file. Behind a security advisory you must acknowledge before any state leaves the perimeter.

Bounded & gated

apply-critique plans a few scoped fixes, runs deterministic checks (build/tests + memory-lint), then summarizes applied vs. rejected. The human gates the result.

Hypothesis, not authority

A clean reviewer can be confidently wrong (one once over-archived live facts). So critique is advisory, gated by deterministic checks + a human — the lesson baked in.

making the ritual happen

No protocol survives a manual step

A ritual that depends on the agent self-triggering fails the moment it lands with an untrained team. So execution got hardened, not just documented — zero manual user step.

Agent-activated triggers

A committed post-commit git hook auto-stubs a missing session log + re-syncs adapters; a CI floor runs memory-lint with zero per-user setup. The tool runs nothing — git & CI do.

First-run init + Windows

One idempotent init.sh regenerates adapters & activates the hook on a fresh clone; .gitattributes pins shell scripts to LF so Git for Windows can't break them.

MERGE.md + safe-write

Conflict resolution that never picks a winner (clash → preserve both + Contradiction); and a read-into-memory-write-once discipline that makes the old truncate-before-read bug structurally impossible.

Enable opens with an exec summary + cancel gate — informed consent before a single file is written.

Vendor support

One shared memory, every major AI tool

"AI enable this repo" detects whatever AI footprint already exists, folds it into the shared layer, and routes that vendor's agent to the one hub — originals preserved, never deleted.

Claude Code Cursor GitHub Copilot Gemini CLI GPT / Codex Cline Roo Code Aider Continue.dev Codeium / Windsurf Zed AI Amazon Kiro Google Antigravity new

Detected & migrated, not re-typed

Steering folds into memory/; chat history becomes session logs; a vendor's skills are promoted into the neutral agent-skills/ layer.

Riding open standards = future-proof

We bet on AGENTS.md (steering) and the open Agent Skills standard (capabilities). Tools that adopt them — Kiro, Copilot, Antigravity, Codex — plug in with little or no glue. Each shipped as a point release.

Principles

Lightweight by design

Vendor-neutral

One shared memory; thin per-vendor pointers route every agent (Claude, Gemini, Cursor…) to one hub. No lock-in.

No-code, markdown

No build/lint/test. Git-native governance: immutable ledger + mutable projection + replay.

Deterministic

Every decision is counting or comparing integers — reproducible across agents and runs.

Guide, don't prescribe

Loop over process; simplicity over completeness. No ceremony.

Never pick a winner

Contradictions and the Vision are surfaced for a human, not resolved silently.

Additive upgrades

Versioned; an older repo upgrades in place via an idempotent, non-destructive ladder.

Proof

It earned its thesis on real work

Node.js → Rust rewrite

A full rewrite delivered against recorded intent — invariants, decisions, and their why pinned and traceable — deterministic, faithful, no drift.

Dogfooded

The tool carries its own memory, Vision, and Blueprint — and the VBDI layer was designed using the VBDI loop. Using the thing to build the thing.

Cross-vendor, real product repo

Driven end-to-end on a large Accenture repo by GitHub Copilot / Gemini 3.1 Pro. It even independently reinvented refresh-metadata — converging on the design that shipped.

An industry-alignment self-assessment's known gaps — supersession, invariant re-checking, contradiction, evaluation, provenance — all shipped as additive, versioned releases.

Differentiators

vs. common practice

DimensionCommon practiceagent-memory
Persistenceper-vendor files / vector storeone shared, git-committed markdown layer
Forgettingsimilarity scores, TTLsdeterministic tiering by counting
Truth maintenanceoverwrite / let stale persistsupersession + contradiction + re-verify
Intent → deliveryunmanagedVBDI trace, grep-detectable drift
Governanceopaquegit history + markdown + human gates
Vendor couplinglocked to one toolneutral; one hub
Capabilitiesper-vendor skill filesneutral agent-skills/: 7 built-ins × 6 adapters; any agent
Ritual executionrelies on the agent self-triggeringgit hook + CI + runnable helpers; arithmetic mechanized, judgment left to the agent; no manual step
Adoption

Point it at a repo

FRESHgenerate memory· MIGRATEfold vendor files in· UPGRADEin place, idempotent
Roadmap & mission

Memory is the deterministic substrate.
The loop is the lightweight control layer.

Next (tracked as Blueprint gaps): greenfield flow · more multi-user hardening (MERGE.md shipped) · an optional SDLC overlay for teams that want it.

Predictable innovation with human partnership — bold ideas, faithful delivery, the human in the loop at every altitude.

agent-memory · v4.26.1
← → / space to navigate · click to advance