Problem
Every AI coding session starts cold. The agent re-reads the repository, re-asks settled questions, and loses decisions between sessions. One answer is more model calls.
What I built
A local-first system that keeps notes, decisions and project history as plain Markdown, indexes them in SQLite, and assembles a bounded context pack when a session starts. One service and command line serve Claude Code, Codex and any MCP client, and each pack also reads Git state and open work items.
Decisions
Search before models. The pack is built from full-text search and rank fusion, so it takes seconds, makes no model calls, and every item in it says why it's there. Markdown stays canonical, so the notes still work without the tool.
Where it stands
Unit, integration, contract and architecture tests gate every pull request. A companion Obsidian plugin ships alongside it. It is early, at version 0.1, with one maintainer.
- Python
- FastAPI
- SQLite FTS5
- MCP
- pytest
- mypy strict
- GitHub Actions