Memoosea dual-path memory system for proactive agents

Source on GitHub

Features

What Memoose does, one paragraph each, with the page that goes deeper.

A knowledge graph with an ontology

Typed entities and facts instead of text blobs. Every fact is one relation between two entities with a one-sentence description, a validity date and an evidence pointer back to where it came from: a file range, a URL, an issue id, or “user said” with a date. The entity types are an ontology you can extend in place or import from OWL, RDF or Turtle. See Tools.

Procedures as a graph

The context an agent most often lacks is not a fact but the next step. Memoose stores procedures in the same graph, after Google’s Procedural Graphs paper: steps as Procedure entities, transitions as relations carrying a condition, an advice and a pitfall. The agent declares its position on a session turn and is handed the transitions two hops out, raw, to decide with. When the session ends with an outcome, every transition it took counts it, so a transition that keeps failing stands out at the next distillation. No model call and no guessing where the agent is. See Vision and the sessions skill.

Recommendation as well as search

recall answers the question the agent thought to ask, across eight retrieval modes routed from the query. A hint before each prompt covers the questions it did not think to ask: memory is searched locally when a prompt arrives and, when something genuinely matches, a short block is put in front of the agent. No model, a few milliseconds, silent when nothing matches. See Automatic memory.

Two ways in, one store

A CLI the agent runs like any other command, memoose recall "who owns billing", and 26 MCP tools for hosts where the shell is restricted. The CLI is the default: the MCP server’s schemas measure about 4,250 tokens resident on every turn, while a command costs nothing until it runs. Both surfaces sit on the same Engine and the same SQLite file. See Tools.

Stays on your machine

No service, no API key, nothing leaves your computer. Each project gets its own memory in one SQLite file, plus a user file for facts that hold in every project; you can inspect, copy or delete any of them. Measured room to roughly a million facts per project before anything needs rethinking. See Configuration.

Changes stay visible

A fact that was true once is never deleted when it changes; it is superseded and stays in history. Disagreements are surfaced as hotspots for a model to judge, every write lands in an append-only provenance ledger, and finished sessions distil into lessons. See Evidence & history.