Lightweight self-improvement loop for AI agents. Capture errors, corrections, and discoveries in a fast one-line format, dedup them, queue recurring or critical lessons for human-approved promotion, and review relevant learnings before major work.