widemem

About widemem

widemem is an Apache-2.0 memory layer for LLM agents built around the premise that an agent which forgets selectively is worse than useless in domains where a wrong recall is expensive. It is local-first: the default deployment is a Python library backed by SQLite and FAISS, with no services to stand up and nothing to page, and it is capable of running air-gapped out of the box. The three properties the project foregrounds are local-first storage, importance scoring and auditability. Importance scoring is the part that distinguishes it from a plain vector store — rather than embedding everything and retrieving by similarity alone, memories carry a scored notion of what matters, so an agent retains the facts it cannot afford to lose rather than whatever happens to be nearest in embedding space. Auditability means the recall path can be inspected after the fact, which is what makes it usable in regulated settings where you have to explain why an agent said what it said. The licensing posture is unusually clean for an open-core product: every tier ships the full Apache-2.0 library with all providers and no gated features, and the paid tiers sell hosting, SLAs and compliance help rather than feature unlocks. Current release is 1.5.0, on Python 3.10 and later.

Key Features

Tags

More AI Agent Infrastructure tools