Engram released: local-first memory for coding agents using SQLite FTS5 and BM25

Engram is a new MIT-licensed open-source project providing local-first memory for coding agents. The system stores all data as plain Markdown files on the local disk, ensuring data portability and simplicity.

Core Architecture

• Search Mechanism: Utilizes BM25 ranking over a SQLite FTS5 index. The index is disposable and rebuilt directly from the Markdown files.
• Hybrid Search: If a local embedding model is already provisioned, cosine similarity results are fused with lexical results using reciprocal rank fusion.
• Offline Recall: The system does not download models at recall time. If no embedding model is present, it defaults to lexical search only, ensuring no network dependency during retrieval.

Testing and Validation

The project includes a test suite that enforces a recall@5 of at least 90% across 20 seeded queries. The author notes that while this small set serves as a regression gate, it is not a comprehensive benchmark.