A skill for Claude Code that gives the agent persistent memory of its mistakes.
The agent maintains a markdown file in your repo (.claude/napkin.md) where it tracks what went wrong, what you corrected, and what worked. It reads the file at session start and writes to it continuously as it works. By session 3-5 the behavior shift is significant — the agent stops making mistakes you've already corrected and starts pre-empting issues before you catch them.
Baby continual learning in a markdown file.
Claude Code
git clone https://github.com/blader/napkin.git ~/.claude/skills/napkinCodex
git clone https://github.com/blader/napkin.git ~/.codex/skills/napkinThat's it. The skill activates every session, unconditionally.
- Session start: Agent reads
.claude/napkin.mdin the current repo. If it doesn't exist, it creates one. - During work: Agent logs mistakes (its own and yours), corrections, patterns, preferences — as they happen, not just at session end.
- Over sessions: The napkin compounds. Session 1 is normal. Session 3 the agent is catching things before you do. Session 5 it's a different tool.
- Agent's own mistakes — wrong assumptions, bad approaches, failed commands
- Your corrections — anything you told it to do differently
- Tool/environment surprises — things about the repo that weren't obvious
- Your preferences — how you like things done
- What worked — successful approaches worth repeating
Lives at .claude/napkin.md in each repo. One per repo. The agent designs the initial structure and adapts it to the project's domain.
You can .gitignore it or commit it — your call. Committing it means other contributors' agents benefit from the same learned patterns. Ignoring it keeps it personal.
MIT