project-mind 0.1.3

Persistent AI context management for multi-tool development — syncs Claude, Gemini, Cursor, Codex and Copilot transcripts into a single LLM-tuning file.

mind

Persistent AI context management for multi-tool development.

Reads conversation transcripts from Claude Code, Gemini, Cursor, Codex, and Copilot after every git commit. Synthesizes them into _mind/mind.md — a token-efficient LLM-tuning file injected into every future session.

Install

pip install project-mind

Setup (per project)

cd your-project
mind init

This creates _mind/, installs a git hook, and writes mind.toml.

Usage

mind sync      # run manually (hook calls this automatically after commits)
mind evolve    # full re-synthesis from all transcripts
mind status    # show tracked tools and last sync time

Supported Tools

Tool	Transcript location
Claude Code	~/.claude/projects/{slug}/*.jsonl
Gemini	~/.gemini/tmp/{project}/chats/*.json
Cursor	~/.cursor/projects/{slug}/agent-transcripts/
Codex	~/.codex/history.jsonl
Copilot	VS Code workspaceStorage/{hash}/state.vscdb

Configuration (mind.toml)

[project]
name = "my-project"

[llm]
provider = "claude"   # claude | gemini | codex

[tools]
enabled = ["claude", "gemini", "cursor", "codex", "copilot"]

[limits]
max_messages_per_sync = 150
mind_max_lines = 150

How it works

1. After every git commit, the hook runs mind sync in the background
2. mind sync scans transcript directories for files newer than last sync
3. Extracts user + assistant text (skips tool calls, thinking blocks)
4. Calls claude -p (or configured LLM) with extracted content + current mind.md
5. LLM rewrites mind.md in place — compressing old entries, adding new ones

mind.md format

## behavior     ← user corrections — never compressed
## context      ← project state — rewritten each sync
## active       ← in-flight tasks
## decisions    ← architectural choices
## lessons      ← what worked / failed
## history      ← compressed timeline