Napkin: Persistent Memory for Claude Code

2026-05-14 · source · source ↗ · raw capture · wiki page

A skill for Claude Code that enables persistent memory across sessions through a per-repository markdown scratchpad (.claude/napkin.md). Allows AI agents to learn from mistakes and improve performance over multiple sessions.

Key Features

Persistent Learning: Agent maintains a markdown file tracking errors, corrections, and successful approaches across sessions
Continuous Logging: Documents mistakes, user corrections, environment discoveries, and preferences as they occur during work
Progressive Improvement: Performance noticeably improves by sessions 3-5, with the agent anticipating issues before correction needed
Per-Repo Design: One napkin file per repository; can be committed (shared with contributors) or ignored (personal use)

What Gets Logged

Agent's own mistakes and false assumptions
User corrections and feedback
Environment-specific surprises
User preferences for task execution
Successful approaches worth repeating

Description

"Baby continual learning in a markdown file." Incremental knowledge accumulation that compounds over time rather than one-time learning.

529 stars, MIT license.