DeepLethe's Utopia puts a bitemporal knowledge graph and reasoning engine behind agent memory, and hits 3,800 stars in a month

DeepLethe has published Utopia, an open-source knowledge management system aimed at AI agents that need durable, auditable memory. The repository was created on 7 August and has picked up 3,839 stars in the four weeks since, with the latest commit at v0.1 as of 4 September. It is licensed Apache 2.0.

What it actually is

Utopia describes itself as an open substrate for knowledge engineering that learns passively and governs itself. In practice it stores facts in a bitemporal graph, which records both when a fact was true in the real world and when the system learned it. That distinction is the whole point: a system that only stores the current version of a fact cannot answer what an agent knew at a decision time, or catch a conflict between an older source and a newer one.

Ingest sources listed in the repository include PDFs, Office documents, web pages, RSS, GitHub, Jira, Notion, S3, and arbitrary HTTP APIs. Search covers full text and vectors together, with inline citations back to the source document, and any OpenAI-compatible chat model can be plugged in on top for question-answering.

What comes with it

Alongside the graph there is a forward-chaining reasoning engine that derives new facts from asserted ones and flags conflicts, five preloaded ontologies (schema.org, W3C Org, PROV-O, FOAF, and IOF Core), role-based permissions for multi-user setups, and an append-only decision ledger for compliance auditing. The whole thing runs as a single Rust binary against a Postgres database with the pgvector extension, and works air-gapped.

Utopia is still at v0.1 and the repository states that migrations only roll forward between schema versions, so anyone experimenting should plan for that.

What it means for a team building software

Agent memory is one of the areas where teams keep reinventing the same partial solution. A vector store answers similarity queries fine but forgets that the fact returned was superseded last week. A relational table remembers the newer version but loses the older one and the timestamp that separated them. A team that wants an agent to reason honestly about what it knew when either builds both and joins them, or writes a scratchpad on top that captures the timeline by hand.

Utopia offers a single store that keeps both timelines, plus a reasoning engine that can act on them and a citation trail back to the original document. The single-binary-plus-Postgres setup is a low-friction shape for a team already running Postgres. The trade is early-version risk: v0.1, no benchmark numbers, no formal comparison against a specific commercial product. A reader with an agent memory problem worth solving now can run it against a subset and see what it does; a reader who needs a stable, warranted knowledge layer for a production system probably wants to watch the release cadence for a few versions first.

Source

• deeplethe/utopia on GitHub