taOSmd
A framework-agnostic AI memory system that keeps every extracted fact tied back to an append-only source archive and verifies it before recall. It runs offline on modest hardware with 8 GB or more RAM.
taOSmd is a memory system for AI agents and assistants. Instead of extracting facts and discarding the original conversation, it stores everything first in an append-only archive that is never edited or deleted, then derives its searchable memory from that archive. A verifier checks each extracted fact against the exact source text and leaves out anything it cannot support, reporting the share of unsupported facts (18.8 percent on the project's own data).
The project measures itself with Recall@5 (the correct evidence session appearing in the top 5 results) and a separate, stricter end-to-end Judge metric that retrieves, generates and grades an answer against a reference. Its shipped configuration reaches 97.0 percent Recall@5, and a factual-recall generator profile lifts the Judge score into the 54 to 61 percent range.
It runs offline on hardware with 8 GB of RAM or more, including a Raspberry Pi, an Intel mini PC or an Orange Pi 5 Plus, needs Python 3.10 or newer, and has no cloud dependencies. It is part of the wider taOS ecosystem.