AI Agent Self-Improvement Log

Problem

When an AI agent participates in coding, debugging, documentation, and tool calls, context often fails to persist. A failed command, user correction, outdated knowledge, or external API error usually exists only in the current session. Later sessions, or agents on another machine, may repeat the same mistake.

How It Works

The skill writes structured learning records into .learnings/ at the project or workspace root, typically across three files:

• LEARNINGS.md: corrections, knowledge gaps, and best practices;
• ERRORS.md: command failures, exceptions, timeouts, and connection errors;
• FEATURE_REQUESTS.md: user-requested capabilities.

Each entry should use a stable ID such as LRN-20250115-001 or ERR-20250115-A3F, plus status, priority, and area tags. When a similar issue appears, search existing entries first, link them with See Also, raise priority if it recurs, and decide whether a systemic fix is needed.

When an experience has cross-file, cross-task, or team-level value, the skill distills it into a short rule and promotes it to project memory such as CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md. In an OpenClaw workspace, it can also be promoted to SOUL.md or TOOLS.md. For recurring patterns from simplify-and-harden, use Pattern-Key for dedupe, track Recurrence-Count, and promote them into system-prompt or project rules after thresholds are met.

Boundaries

The skill fits agent workflows that maintain workspace files and want scattered corrections to become stable prompt guidance. Avoid storing secrets, tokens, full configuration files, or long raw output in logs; short summaries, redacted excerpts, and reproducible steps are preferred. Hooks and cross-session sharing are optional and should be enabled only in trusted environments.