Triage Stale Memory

CLAUDE.md is the doctrine. This card is the executable procedure for one specific situation: a session did real work in a memory-hook log but never recorded a lesson or decision for it, so amplify_context_load ends its output with a stale-memory warning.

Stale memory is the second shape of memory drift (the first is hallucinated write-success). Treat it the same way: nothing counts until Amplifier reads it back.

When to use

• amplify_context_load output ends with a stale-memory warning, OR
• you suspect a previous session logged work to ~/.claude/memory/<date>.md

but forgot to persist any lesson/decision.

Procedure

1. Audit. Run amplify_audit_freshness to get the exact list of stale

memory days (files newer than the latest Amplifier write).

1. Draft, don't record. For each stale file, run

amplify_promote_from_memory_md({ memory_file: "<path>" }). This records nothing — it returns DRAFT suggestions only.

1. Classify each draft into exactly one of:

• decision — an architectural / tooling / process choice → record with

amplify_decisions({ op: "track", ... }).

• lesson — a mistake, insight, or confirmed fix → record with

amplify_learn(...) (or amplify_record_claim if it is an unverified inference — see record-verified-lesson).

• discard — noise, one-off, or already captured. Do nothing.

1. Record only load-bearing items. A draft is load-bearing only if a future

session would make a worse decision without it. When unsure, prefer amplify_record_claim (a claim weighs 5× less than a confirmed lesson) over inventing a confident lesson.

1. Read-back gate. After each write, confirm Amplifier actually persisted it

(the tool re-reads the row; a real write returns a numeric id, a failure returns ERROR: ... NOT recorded). Never claim the stale work was captured until read-back confirms it. If you see an ERROR, switch to investigate-write-failure — do not retry blindly.

Required output artifact

Write a triage record to:

./knowledge/stale_memory_triage_{YYYY-MM-DD}.md

Use today's date (run date first — do not assume the date). The file must list, per stale file: each draft, its classification (decision/lesson/discard), and — for recorded items — the returned numeric id (proof of read-back). Discards get a one-line reason.

Anti-patterns

• ❌ Recording every draft "to be safe" — this floods the Pattern Oracle with

noise and dilutes real signal.

• ❌ Reporting "captured the stale day" before any read-back returned an id.
• ❌ Promoting an unverified inference as a confirmed lesson. Use a claim.