Captures learnings, errors, and corrections from every session to enable continuous self-improvement for your AI coding agent.
by
pskoett
self-improving-agent records errors, corrections, missing capabilities, and better approaches in a structured .learnings/ folder.
It turns recurring issues into durable project memory by logging consistently and promoting proven learnings into project instruction files.
It activates on command failures, user corrections, missing features, API errors, outdated knowledge, or better approaches.
Events are appended to .learnings/ files with timestamps, category, context, and resolution notes.
Before major tasks, the agent reviews prior learnings to avoid repeating mistakes.
Confirmed learnings are promoted into project memory files like CLAUDE.md or AGENTS.md.
Records unexpected command failures, tool errors, and API failures to .learnings/ERRORS.md.
When a user says "No, that's wrong" or "Actually...", the correction is logged under the correction category.
Logs missing capabilities to .learnings/FEATURE_REQUESTS.md for later planning.
Identifies outdated or incorrect knowledge and logs it with category knowledge_gap.
Stores better approaches for recurring tasks under category best_practice.
All learnings are stored as human-readable Markdown for easy review and version control.
Supports OpenClaw workspace injection and inter-session tools for sharing learnings.
Prevents the agent from repeating the same mistakes session after session.
Share the .learnings/ folder to accumulate collective agent knowledge.
Find the root cause of recurring errors by reviewing structured error logs.
Improve agent behavior over time without manual prompt engineering.
Use feature requests to prioritize tooling improvements.
npx clawhub@latest install self-improving-agentRun in your terminal, or click the Install button at the top of this page for one-click setup.
In the .learnings/ directory using LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md.
No. It runs locally and only writes to markdown files.
Confirmed items are promoted into project instruction files like CLAUDE.md, AGENTS.md, TOOLS.md, or SOUL.md.
Any agent that reads markdown instructions can use it, including Claude Code, Cursor, and Windsurf.
Failures, user corrections, missing features, tool/API errors, knowledge gaps, and better approaches.
Entries include a timestamp, category, summary, context, and suggested resolution in a consistent template.
Yes. The skill supports an optional OpenClaw hook for session-start reminders.