agentsclimarketplace
Sign in

Memory audit

Skill Wanyaldee/wanyaldee-skills/skills/memory-audit

One-shot audit of all persistent auto-memory directories (~/.claude/projects/*/memory/) — index gaps, repo-duplicating dumps, stale or contradicted facts, format violations, lost provenance. Use when the user says "/memory-audit", "audit memories", "メモリ監査", "check my memories", or asks what's stale or wrong in memory. Report-only by default; --fix applies mechanical repairs only.From its SKILL.md

Install
npx -y skills add Wanyaldee/wanyaldee-skills --skill memory-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

2.9 KB, as published. Nobody here has run it

Memory audit

Sweep every project's memory directory, grade each file against [[memory-discipline]], and report — most severe first. Do not modify anything unless invoked with --fix, and even then apply only the mechanical repairs listed at the bottom.

Procedure

  1. Enumerate. ls -d ~/.claude/projects/*/memory — for each directory, read MEMORY.md and every other *.md in full. They are small; read all of them, don't sample.

  2. Index integrity. Diff the file list against the index: every memory file needs exactly one MEMORY.md line, and every index line a file. Unindexed = never recalled; dangling = recall promising a 404.

  3. Attribution. For each file, note origin: if present. If only originSessionId exists, try to resolve the model: grep -o '"model":"[^"]*"' ~/.claude/projects/*/<sessionId>.jsonl | sort | uniq -c (quote the glob path carefully — project dirs start with -, so prefix ./ or -- when globbing). A purged transcript = lost provenance; flag it, and record the resolution while it's still possible.

  4. Per-file checks, in severity order:

    • Contradicted: a claim you can cheaply verify against current state (a repo's visibility, a file's existence, a version, a "pending" decision since resolved) that is now false.
    • Repo duplication: body restates a document the project already has (the giveaway: it links to its own source-of-truth). Should be pointer
      • delta.
    • Unindexed / dangling (from step 2).
    • Mixed volatility: durable setup facts and current-status in one file — flag the split point.
    • Format: missing Why/How on feedback/project types, non-kebab name, type outside metadata:, relative dates, missing origin.

  5. Report. Group by project, one line per finding: file, defect, the fix. Lead with a one-paragraph verdict (worst finding + overall health). Note per-model patterns only if attribution actually shows one. End with the fix list split into mechanical vs needs the user's judgment.

--fix (mechanical repairs only)

Allowed: add missing index lines; remove dangling index lines; backfill origin: where step 3 resolved it; convert relative dates when context pins the absolute one. Everything else — deleting stale memories, shrinking dumps to pointers, splitting files — is judgment: list it, don't do it. Never touch memory content under --fix beyond frontmatter and dates.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.