Agent Self-Improvement Logging

Problem It Solves

Agents often repeat the same coding mistakes: user corrections are not retained, command failures are patched temporarily, and missing capabilities disappear after the session. If these lessons stay only in chat transcripts, the next session may hit the same traps. This skill turns corrections, errors, and feature requests into Markdown logs in the project, giving later coding agents a searchable, repairable, promotable knowledge base.

How It Works

The core loop is log, archive, and promote:
- Initialization: create .learnings/ in the project root with LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md, without overwriting existing files.
- Immediate logging: when users correct behavior, commands fail, knowledge is outdated, or external tools fail, append typed entries using IDs like LRN-YYYYMMDD-XXX, ERR-..., and FEAT-... for easy lookup.
- Status tracking: move entries from pending to in_progress, resolved, wont_fix, or promoted, and add resolution notes or related-entry links.
- Project memory promotion: when a lesson applies across files or tasks, distill it into a short rule and write it to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; OpenClaw workspaces may use SOUL.md or TOOLS.md.
- Recurring pattern handling: search existing logs first, then add See Also, raise priority, or promote a prevention rule once the pattern meets recurrence conditions.

Boundaries

It fits engineering-style agent workflows with persistent project files and cross-session context. It is less useful for one-off scripts or ephemeral tasks without a durable workspace. Logs should avoid secrets, tokens, environment variables, and full source dumps, preferring short summaries and redacted excerpts. Promoted files should stay concise: reusable rules, not long incident reports.