← ClaudeAtlas

memory-auditlisted

Audit Claude's persistent auto-memory (MEMORY.md + memory files) against current workspace state; propose stale-entry updates, merges, and deletions for user approval, then apply the approved batch. Use periodically (e.g. when MEMORY.md grows past ~30 entries or quarterly) or when recalled memories keep contradicting the repo.
hsigstad/research-kit · ★ 0 · Data & Documents · score 73
Install: claude install-skill hsigstad/research-kit
# /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 Entries rot in three ways — the fact changes (a bug gets fixed, a file moves, a decision is superseded), the fact gets absorbed into durable docs (rules, findings, decisions logs), or two entries accumulate that cover the same thing.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 lar