Claude Memory Audit

Audits Claude Code conversation history to find repeated errors and anti-patterns, then writes project-scoped memory files so future sessions don't repeat the same mistakes.

Runs every 6 hours, processes only unseen conversations (cursor-tracked), and git-commits all memory file changes.

Usage

./start.sh                # Start UI + audit loop in tmux
bash run-audit.sh          # Single audit run
bun run server.ts          # UI only on http://localhost:3100

How it works

1. extract-unseen.sh reads ~/.claude/projects/ JSONL logs, skipping already-processed lines (tracked in cursor.json)
2. Unseen conversation chunks are fed to Claude via run-audit.sh with the analysis prompt
3. Claude identifies repeated failures, recurring errors, and anti-patterns
4. write-memories.py writes memory files to the correct project's ~/.claude/projects/<project>/memory/ directory
5. Changes are git-committed with an audit summary
6. Results are displayed in the web dashboard

Files

• audit-prompt.md — Prompt instructing Claude what patterns to look for
• run-audit.sh — Orchestrates: extract -> analyze -> write memories -> commit
• extract-unseen.sh — Reads JSONL logs, outputs only unprocessed lines
• update-cursor.sh — Marks all current JSONL lines as processed
• write-memories.py — Parses audit JSON, writes memory files to project dirs
• cursor.json — Tracks last-processed line per conversation file (git-tracked)
• memory-log.md — Append-only log of all memories created (git-tracked)
• reports/ — JSON audit reports (gitignored)
• server.ts / frontend.ts — Dashboard UI