memory-audit

---

name: memory-audit description: Audit Claude's per-user memory for this project and produce a report. Asks at the start whether the user wants a non-technical summary or a full technical audit. Never edits memory — produces a report only. Usage: /memory-audit allowed-tools: Read, Grep, Glob, Bash, AskUserQuestion

You are running a memory audit for the user's per-user Claude memory in this project.

Hard rule: this skill produces a report only. Do not edit, rename, move, or delete any memory file. If the user wants to act on findings, that is a separate follow-up turn after they review the report.

This skill focuses on the internal health of the memory directory itself. For how memory interacts with the wider repo — conflicts with CLAUDE.md, rules that never reach subagent prompts, broken doc links — see /context-audit.

Step 1 — Ask the user which report style they want

Before reading anything, call AskUserQuestion with:

• question: "What kind of memory audit would you like?"
• header: "Report style"
• multiSelect: false
• options:
  • label: "Non-technical summary" description: "Plain-English overview for project management or non-engineering review. Each remembered item is described in everyday language with a clear recommendation."
  • label: "Technical audit" description: "Full engineering audit: classification, stale-reference checks, orphaned files, and structural smells. For engineers maintaining the memory."

Wait for the answer before continuing. The user's choice determines the report format in Step 5.

Step 2 — Resolve the memory directory

The memory directory lives at ~/.claude/projects/<encoded-cwd>/memory/, where the encoded form replaces drive colons and path separators with -.

Reliable way to find it: list ~/.claude/projects/ and pick the directory whose decoded name corresponds to the current working directory. On Windows the path will look like C:\Users\<user>\.claude\projects\C--Users-<user>-...\memory\.

If the directory does not exist, stop and tell the user: "No memory has been recorded for this project yet — there is nothing to audit."

Step 3 — Read everything

• Read MEMORY.md.
• Read every .md file in the directory — both those the index points to and any that are not indexed. Do not assess a memory you have not read.
• For each memory, note its declared type from frontmatter (one of: user / feedback / project / reference).
• List the directory contents and identify:
  • Files in the directory not indexed in MEMORY.md
  • Index entries pointing to files that don't exist

Step 4 — Verify references against current code

For any memory entry that names a specific file path, class, function, or directory in the codebase, read the current code to check the reference still resolves. If a class has been renamed, a file deleted, or a path moved, treat that entry as a candidate for refresh or removal.

Do not guess. If the entry says "the XyzService does …", actually search the code for XyzService.

Step 5 — Produce the report

If the user picked "Non-technical summary"

Write for a non-engineer reading this for project-management purposes. No jargon, no file paths, no markdown class names in backticks.

For each memory, produce a short block in this shape:

#: Memory number Memory: One plain-English sentence describing the rule or fact. Why it was added: The original reason, in everyday language. If the entry says it was added after a specific incident, say so simply ("Claude was told this after it ran a series of slow git commands the user had to approve one by one"). Still useful?: One of: Yes, Probably, Maybe not, No longer. Add one short sentence explaining your judgment. Recommendation: One of: Keep / Discuss with the team / Remove. ────────────────────────────────────────

After all entries, write a short overall summary in 3–5 sentences:

• How many memories there are
• Whether the set looks healthy or cluttered
• Any patterns worth raising with the team (for example: a lot of "don't do X" rules, or several overlapping rules covering the same situation)

End the report with this exact line on its own:

Want the full engineering breakdown — code verification, classification rules, and structural analysis? Run /memory-audit again and choose "Technical audit".

If the user picked "Technical audit"

Produce a full audit. For each memory entry, classify it as exactly one of:

• Keep — durable, applies broadly across tasks, not derivable from code or docs
• Move — real rule but situational; should live in docs/, in a .claude/agents/<name>.md subagent prompt, or in CLAUDE.md rather than in always-on memory
• Refresh — references files/classes/paths that have rotted; either update the entry or flag the discrepancy
• Merge — overlaps another entry and they should be combined
• Drop — situational one-off, no general principle, low signal-to-noise

For each entry include:

• Its name and file
• The classification with one-sentence rationale
• For any code reference: whether it still resolves (verified by reading the code)

After the per-entry list, report:

• Orphans: files in the memory dir not in MEMORY.md, and index entries pointing to missing files
• Type sanity: does each entry's body match its declared type? A reference entry should point to a resource that still resolves (verify the URL or path); a project entry should use absolute dates, not "last week" / "recently"; a feedback entry should carry both a why and a how to apply. Flag mismatches.
• Prohibition ratio: count of entries whose rule is a "don't / never / avoid / stop" against total entries. If above ~60%, call this out as a memory-as-corrections-log smell — memory is meant to shape behavior, not log every past mistake.
• Justification quality: entries whose "Why" is a single past incident with no general principle — usually candidates for Move or Drop.
• Topic clustering: if several entries pile into one narrow workflow (e.g. one investigation), suggest collapsing them into a single doc or subagent prompt.

If any of the memories are classified not as "Keep", end the technical report with a prioritised action list: which entries to address first, and whether the recommended next step is editing the entry, moving it elsewhere, or dropping it.

After the report

Stop after delivering the report. Do not begin editing memory. If the user asks you to act on findings, treat each edit as a separate operation and confirm before writing.