/memory-audit — Prune and refresh Claude's persistent memory

The memory counterpart of /distill-style. Auto-memory is an append-mostly store: entries are written when something non-obvious is learned, but nothing verifies them later. Entries rot in three ways — the fact changes (a bug gets fixed, a file moves, a decision is superseded), the fact gets absorbed into durable docs (rules, findings, decisions logs), or two entries accumulate that cover the same thing. Rotten memories are worse than none: they are recalled with the authority of "what I know" and contradict the repo.

Arguments

• /memory-audit — audit the current session's memory directory (the one whose MEMORY.md loads into context).
• /memory-audit --all — audit every store under ~/.claude/projects/*/memory/.
• /memory-audit --dry — produce the report only; skip the apply step even if the user approves inline.

Workflow

1. Inventory

Read MEMORY.md and every memory file it indexes. Also list memory files NOT indexed in MEMORY.md (orphans) and index lines whose file is missing (dead pointers) — both are findings by themselves.

2. Verify each entry against the workspace

For each memory file, extract its checkable claims: file paths, flags, numbers, states ("X is pending", "Y is broken", "use Z workaround"). Check them against the current repo state — read the referenced files, grep for the flags, check dates. Delegate this to cheap subagents in batches when the store is large; each verifier returns entry → {holds | changed | absorbed | unverifiable} with one-line evidence. Classify:

• current — claims hold; keep unchanged.
• stale — a checkable claim no longer holds (file moved, bug fixed, number superseded). Propose an update or deletion.
• absorbed — the content now lives in a durable artifact (a rule file, decisions.md, findings.md, a method note). Propose deletion, citing where it lives now. Memory duplicating docs is recall noise.
• duplicate/overlap — two entries cover one fact. Propose a merge into the better-named file.
• unverifiable — states a preference or event the repo cannot confirm (e.g. user preferences, conversations). Keep unless the user says otherwise; flag only if it contradicts another entry.

Session-scoped trivia that should never have been saved (one-off task state, things fully derivable from the repo) is proposed for deletion regardless of freshness.

3. Write the proposal report

One numbered proposal (M1, M2, …) per action, grouped by action type (delete / update / merge / index-repair). Each proposal shows the entry's name, its current one-line description, the evidence (what was checked, what came back), and — for updates — the exact replacement text. Put the report in the conversation directly if it is short (≤ ~15 proposals); otherwise write it to the scratchpad and give the path.

4. User approval

Same protocol as /distill-style: the user replies "approve all", "approve M1, M3", "approve all except M2", "reject M4 — <reason>". Never apply without approval.

5. Apply

For approved proposals: edit or delete the memory files, update MEMORY.md index lines to match (every file change has a matching index change), fix [[...]] links that pointed at deleted/renamed entries. Deletion is real deletion — memory is a working cache, not history; the git-tracked repos are the archive.

6. Final report

Entries kept / updated / merged / deleted; store size before and after; any recurring rot pattern worth a note (e.g. "session-state entries keep appearing — stop saving those").

Rules and gotchas

• Never auto-apply. Every mutation requires explicit approval.
• Bias toward deletion. The index is read every session; each line must earn its context cost. When in doubt between "update" and "delete because the repo now records it", delete.
• Don't churn user/feedback entries on repo evidence alone — preferences aren't falsified by code. Flag contradictions between them instead.
• Keep name slugs stable on update so inbound [[links]] survive.
• --all mode: audit each store against its own project root, not the workspace root.
• Cadence: worth running when MEMORY.md passes ~30 lines, after a big refactor/migration, or when a recalled memory is caught wrong twice in one week.