The Un-Brain — Graphnosis — is a local-first knowledge multi-graph for private AI memory, reached by AI clients over the Model Context Protocol. Its contribution is a synthesis rather than any single first: a co-equal dual graph (directed and undirected edges as two independently-queryable classes over the same nodes), a total, deterministic adjudication contract (every detected conflict recorded in an inspectable lane and surfaced for the owner to adjudicate, none silently resolved — with no language model in the loop), and an inspectable deterministic substrate that is architecturally indelible (node-set indelibility), converges under an order-independent merge (last-writer-wins determinism), retrieves hub-independently, and is constructed with zero embedding or generation API calls. The adjudication contract is shared with concurrent 2026 work; what a source audit of the 2026 wave did not find is the three properties held together in one owner-held, local-first store. On the public LongMemEval benchmark the zero-embedding-API configuration scores 64.6%, a fully on-device configuration 41.6%, and a cloud-paired configuration 78.0%; a controlled retrieval ablation (answer model fixed) attributes +13.2 points to the typed-graph structure alone.
Nelu Lazar (Thu,) studied this question.