Automatically capture errors, corrections, and best practices to evolve agent capabilities safely over time.

• Records seven types of triggers (failures, user corrections, feature requests, API errors, knowledge gaps, better approaches, task reviews) into structured markdown files with deduplication to avoid low-value entries
• Promotes repeated patterns (≥3 occurrences) from learning logs into permanent capability files (AGENTS.md, TOOLS.md, SOUL.md) or extracts them as independent skills
• Maintains a JSONL changelog tracking all operations (add, promote, extract, resolve) with timestamps and metadata for auditability and pattern analysis
• Enforces safety guardrails (ADL anti-drift protocol, VFM value-first scoring) to prevent capability drift and prioritizes stability over novelty