AI Agent Self-Improvement Learning Log

Problem

Agents often hit one-off failures while working on code: failed commands, user corrections, changing API behavior, outdated docs, missing capabilities. If that context stays only in a chat session, the next run can repeat the same mistake, and other agents or contributors cannot reuse it.

How It Works

This skill turns feedback into structured Markdown logs, usually under .learnings/:

• ERRORS.md: command failures, exceptions, external tool errors, with reproducible details and output.
• LEARNINGS.md: corrections, knowledge gaps, and best practices, optionally categorized as correction, knowledge_gap, or best_practice.
• FEATURE_REQUESTS.md: user-requested capabilities.

Entries use IDs such as LRN-YYYYMMDD-XXX, ERR-YYYYMMDD-XXX, and FEAT-YYYYMMDD-XXX, with statuses like pending, in_progress, resolved, and promoted. When a lesson recurs and applies across tasks, it can be distilled into a short rule and written into CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in an OpenClaw workspace it may map to SOUL.md or TOOLS.md.

Boundaries

It is not an automatic fixer and does not directly patch code. Its value is a searchable, promotable feedback loop. If the project wants team-wide reuse, decide whether .learnings/ should be tracked in Git; if logs are never reviewed, they become noise. It suits long-running agent workflows, not one-shot task execution.