Memory Audit

Memory Audit

Agents now remember. Nothing checks what they remember. A memory line written on day one is still steering the agent on day 365, and a note that was true in March makes the agent confidently wrong in July. Every entry in a memory file is a claim about the repo. Claims age. This skill fact-checks them, line by line, with receipts.

What Counts As Memory

Audit every file the agent loads without being asked:

• CLAUDE.md at repo root and parents, plus anything it imports
• AGENTS.md, GEMINI.md, .github/copilot-instructions.md
• .cursor/rules/, .cursorrules, .windsurfrules
• the agent's per-project memory directory (for Claude Code, the memory/ folder under the project's config dir, MEMORY.md first)

Inventory them with line counts before judging anything. Memory nobody can list is memory nobody is auditing.

Method

1. Inventory. List every memory file found, its size, and its last modified date from git.
2. Extract claims. Split each file into atomic claims: a path, a command, a flag, a version, a rule, a recorded decision. One claim, one row.
3. Verify each claim, cheapest check first. Does the path exist? Is the command still in the Makefile or package.json? Is the flag still in the config? Does the version match the lockfile? Does the named function still exist (grep)? For dated facts, check git log around the date, not memory of the date.
4. Cross-check the files against each other. Two instruction files that disagree are worse than one stale file: the agent obeys whichever it read last.
5. Report the table. Never fix anything.

Verdicts