Memory Correction Loop

The AI kept repeating mistakes I had already corrected. My own project instructions even say "James has corrected this repeatedly." Written rules also bloated. One team guardrails file reached 453 lines against a 210-line ceiling, and one written rule was broken in 14 sessions.

1. When I correct the AI, it saves a small memory: the rule, why (the real incident), and how to apply it.
2. A one-line pointer goes into an index that loads every session, kept under 200 lines.
3. On the team toolkit, every build retro files its findings to an inbox, and a triage skill folds them in with one PR.
4. A ledger tags each finding as new, recurred or covered. Lessons that recur in 2+ reports become hooks or CI checks.
5. A CI ratchet caps knowledge-file length, so files cannot grow past their cap.

• The AI stops repeating mistakes it has already been corrected on, and repeats keep falling: from 25% to 12% on our team toolkit, August to October.
• A lesson one person learns reaches the whole team through the shared toolkit.