← sources

Engram: local-first memory for coding agents. SQLite FTS5 + BM25, optional local embeddings, no network at recall time (MIT)

source: reddit (reddit:LocalLLaMA)published 2026-10-03 13:28ingested 2026-10-04 02:38

retained text

I'm the author. Engram is free and MIT licensed.

Everything is plain Markdown on your disk. Search is BM25 over a SQLite FTS5 index that is rebuilt from the Markdown, so the index is disposable. If a local embedding model is already provisioned, cosine results are fused with the lexical ones by reciprocal rank fusion. Recall never downloads a model, so with no model present it simply stays lexical.

The test suite enforces recall@5 of at least 90% across 20 seeded queries. That is a small set, so it works as a regression gate, not a benchmark. Walkthrough video above. Repo: [https://github.com/utsapoddar/engram](https://github.com/utsapoddar/engram)

document fetched 2026-10-04 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1wwna5h/engram_localfirst_memory_for_coding_agents_sqlite