Review Memory¶

The review is the manage step of the memory loop: it replays the ledger, re-tiers facts, archives the faded ones, applies supersessions, prompts invariant re-verification, and scans for contradictions and altitude drift.

When it runs¶

On cadence — at least every review_every sessions (default 10).
On demand — any time you ask.
When memory-lint raises a [review-overdue] or [continuity-bloat] advisory, so a lapsed review can't silently hide.

Ask the agent¶

"Run the memory review."

What happens¶

flowchart TD
  A[Replay the ledger<br/>count sessions_since_last_used] --> B[Re-tier facts<br/>refresh-metadata]
  B --> C[Decide what to retire<br/><i>agent judgment</i>]
  C --> D[Move to archive<br/>archive-fact]
  D --> E[Apply supersessions]
  E --> F[Prompt invariant re-verify]
  F --> G[Scan for contradictions / drift]
  G --> H[memory-lint: verify clean]

The split is deliberate — mechanize the arithmetic, leave the judgment to the agent:

Step	Who does it
Recompute tier / `uses` / `last_used`	`refresh-metadata` (deterministic)
Decide which faded facts to retire	the agent (judgment — never automated)
Perform the archive move	`archive-fact` (truncation-proof)
Verify the result	`memory-lint` (read-only)

Safety¶

Never truncate a memory file when scripting a move. Append with >> / append-mode for the archive + INDEX.md; read-into-a-variable-then-write for continuity.md.
Run memory-lint after any scripted memory mutation. It catches a truncation (count drops, links dangle); git-tracked files recover via git checkout HEAD -- <file>.
Archived ids go in the ## Memory Review block of the session log, never under ## Memory References (that would re-arm the over-archival guard and falsely reactivate the fact).

For the authoritative ritual, see REVIEW.md and DECAY.md.