Reflective Learning Loop

What You Get A complete SKILL.md that teaches your OpenClaw agent to run a structured self-reflection after every complex task, filing: • ERROR_LOG.md — mistakes logged with root cause, solution, and a one-line prevention rule • obsidian/Knowledge/ — reusable patterns extracted and saved in structured format • _skill-index.md — a running index of everything your agent has learned, one line per entry • MEMORY.md — behavior-changing lessons that update your agent's long-term operating rules

How It Works

1. Agent completes any task involving 3+ tool calls, debugging, or a corrected mistake
2. Qualifies first — simple lookups and single-step answers are skipped automatically
3. If a mistake was made, logs it to ERROR_LOG.md before anything else
4. If a new reusable pattern emerged, writes it to your knowledge vault with approach steps and pitfalls
5. Silent by default — no narration, no summaries, doesn't interrupt your workflow

Why This Exists Every session, your agent starts from scratch. It has your memory files but it doesn't carry forward the small procedural lessons — the API endpoint that needs a specific header, the routing rule with the exception nobody documented, the debugging approach that cut through a hard problem last time.

Without a learning loop, your agent relearns those lessons every time. Sometimes it makes the same mistake twice. You correct it. Next session it's gone again.

With this running, every hard-won fix gets preserved. After 30 days, your agent handles recurring tasks with fluency that only comes from accumulated experience — because it actually has accumulated experience.

Customization Works with or without Obsidian — all files are plain markdown. Adjust the qualification threshold, change the vault path, add custom log formats. The structure is simple enough that your agent can extend it on its own.