Managing Agent Memory

Reviewer and maintainer agents write notes to .claude/agent-memory/<agent>/. Whether those belong in the repository is a per-project decision — some projects want them shared, others deliberately do not — and whether the existing ones still deserve to be there is a judgement call that has to be re-made every time, because notes rot: the code they describe gets renamed, and the bugs they warn about get fixed.

Three failure modes this skill exists to prevent:

1. Asking the same question forever. Re-prompting on every commit is noise.
2. Answering it once for everything. A decision taken in one project must never silently apply to the next one. There is no global policy.
3. Committing notes that are now WRONG. A note claiming an open security hole that has since been closed, or naming a symbol a rename deleted, is worse than no note — it sends the next agent down a path that no longer exists and gets believed because it looks specific.