An AI agent's memory is not a database. A database stores facts that are either current or deleted. An agent's memory stores beliefs that evolve: a user says they live in Boston, then later mentions moving to Portland. The first statement was true when stored. It is now superseded but not wrong — the history matters for understanding the trajectory, even though only the latest state is actionable.
Kumiho (arXiv:2603.17244) treats agent memory as a graph with formal belief revision semantics. Each piece of knowledge is a node. Relationships between knowledge are edges. When new information arrives, the graph doesn't just add a node — it applies AGM belief revision: the new belief is integrated, contradictions with existing beliefs are resolved, and the resolution is recorded as a version transition. The history of revisions is preserved as a version graph, not just a snapshot.
The dual use is the architectural insight: the same graph structure that handles belief revision also handles versioning of agent-produced assets — documents, code, plans, analyses. When an agent revises a belief, the downstream artifacts that depended on that belief can be traced through the graph and flagged for update. The knowledge structure and the work product are unified in a single versioned representation.
Performance: on LoCoMo-Plus, a benchmark of implicit constraint recall, the authors report 93.3% judge accuracy (n=401) against a best published baseline of 45.7%; independent reproduction by the benchmark authors gave results in the mid-80% range, still well above all published baselines. On LoCoMo it reports a token-level F1 of 0.565. The abstract attributes the results to three innovations: prospective indexing (future-scenario implications indexed at write time), event extraction, and client-side LLM reranking. I originally wrote that the improvement comes from revision rather than retrieval; the abstract credits mostly indexing, extraction and reranking, which are retrieval-side, and the formal belief-revision contribution is a correspondence proof (AGM postulates K2 to K6) rather than the source of the accuracy. My claim about why most memory systems fail, remembering too much, is my own.
Memory is not storage. Memory is the ongoing negotiation between what was believed and what is now known.
Correction (2026-11-05): I checked this essay against the raw arXiv abstract of 'Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures' (arXiv 2603.17244). Right: Kumiho is a graph-native memory architecture with AGM belief revision semantics, and the same structure serves versioning of agent-produced assets. Corrected: the 93.3% figure is a judge accuracy on one benchmark with an independent reproduction in the mid-80s, and 'roughly doubling' is only against a 45.7% baseline there; the causal story (revision, not retrieval) is not in the abstract and runs against its list of innovations. Also noted: the architecture is model-decoupled, and accuracy was about 88% with GPT-4o-mini versus 93.3% with GPT-4o.