
- Category
- AI
- Website
- github.com
- Language
- English
- Listed
- 10d ago
- Stars
- 278
- Forks
- 15
timgordontg/engrim ↗, Python, MIT, Last commit 2d ago
engrim keeps a project's decisions in a local SQLite file instead of the conversation, so moving between Google Antigravity, Claude Code, Cursor, Windsurf and the Codex CLI on the same repository does not start from nothing. Memory is scoped per project and never leaves the machine. Retrieval runs SQLite FTS5 keyword search and static model2vec embeddings separately, then fuses the two rankings, and what comes back is a curated pack of roughly four thousand characters rather than a transcript replay. The project reports a 105-session case study on a trading codebase in which about 153,000 tokens of accumulated work reloaded as under a thousand. Installation is pip install engrim, and engrim setup looks for whichever agent directories exist on the machine and wires hooks or an MCP server into each.