Self-Improvement Memory

Problem

AI coding agents often fail to retain context: command failures, user corrections, stale knowledge, and repeated patterns remain trapped in a single session, so later sessions repeat the same mistakes. This skill writes those signals to structured Markdown files so experience can outlive the conversation.

How It Works

The core is a .learnings/ directory with LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md. When a command fails, the user corrects the agent, an API behaves unexpectedly, or a capability is requested, append entries with ID, Status, Priority, and See Also fields. If a new issue resembles an existing one, search first, link entries, and raise priority when it recurs. When a lesson is reusable across tasks, distill it into a short rule and write it to CLAUDE.md, AGENTS.md, .github/copilot-instructions.md, or OpenClaw's SOUL.md and TOOLS.md, then mark the source entry as promoted.

Scope and Caveats

This is best for repositories where agents need consistent conventions across sessions. It should not be used to store secrets, tokens, private keys, environment variables, or full raw output; prefer summaries, redacted excerpts, and reproducible steps. Teams must also decide on a .gitignore policy explicitly, because learned entries can be useful shared context or local noise depending on the workflow.