agent-memory-governance

Audit persistent agent memory by validating stored facts against live git and gh output, classifying entries as stale, duplicate, or unverifiable, and proposing per-entry corrections with provenance. This stops stale or unsafe facts from steering future sessions.

Memory governance verifies recalled repository state by checking remembered PR, branch, and repo conditions against live git and gh output. If the live state differs, entries are marked stale or wrong-when-written and corrected replacements are proposed.

Clean up secret-leaking or forbidden content in an agent memory store by classifying entries as forbidden-content and proposing delete dispositions. The process requires human-approved apply steps before any edits are made to ensure safe removal.

Yes, you can use git and gh verification to check if a remembered PR is still open. The skill queries live GitHub state, compares it against the memory entry, and marks the fact stale if the PR was already merged, proposing a corrected replacement.

Yes, human approval is required before agent memory corrections are applied. The skill proposes per-entry dispositions such as keep, correct, merge, or delete, but stops for approval before editing files, ensuring no unsafe or unverified changes are committed automatically.

Agent memory becomes dangerous without provenance and absolute dating because stale, duplicated, or unverified facts can steer future sessions incorrectly. Requiring absolute-dated provenance and one-fact-per-entry hygiene ensures that recalled context remains trustworthy and auditable.