Self-Improving Agent Memory

Problem

Coding agents often hit command failures, API exceptions, user corrections, or tool limits during longer tasks. Those signals usually exist only in the current context, and once the session ends they can disappear. In team workflows, error causes, project conventions, and tool quirks are hard to preserve through informal handoff.

How It Works

The skill writes these signals into Markdown files under .learnings/, so later agents can search, update, and promote them:
- ERRORS.md: non-zero exits, stack traces, timeouts, API or external tool failures
- LEARNINGS.md: user corrections, knowledge gaps, and safer approaches
- FEATURE_REQUESTS.md: capabilities users want

Entries use TYPE-YYYYMMDD-XXX IDs and can include Status, priority, Area, and See Also. If a learning only fixes one issue, it can remain in the log. If it applies across files or tasks, it is distilled into a short rule and added to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md. OpenClaw workspaces can also use SOUL.md and TOOLS.md for behavioral conventions and tool notes.

Fit and Caveats

It is better suited to ongoing development, agent collaboration, or repositories with long-term tasks than to one-off chat. Avoid logging secrets, tokens, environment variables, full source files, or raw command output; prefer redacted summaries. Cross-session reads and sends are optional, and should only happen in trusted environments when the user explicitly asks for them. Hook reminders add a small context cost and should be enabled as needed.