🤖 AI Summary
This work addresses the lack of a unified architecture and formal semantic foundation in existing AI agent memory systems, which struggle to jointly support cognitive memory and versionable asset management. The authors propose Kumiho, a graph-native cognitive memory architecture that establishes a formal correspondence between AGM belief revision theory and attributed graph operations, yielding a unified memory model supporting immutable revisions, typed dependencies, and URI-based addressing. The architecture introduces three key innovations: prospective indexing, event extraction, and client-side LLM reranking. It employs a dual-storage design using Redis and Neo4j, integrating full-text and vector retrieval to enable decoupled hybrid retrieval and reasoning. Evaluated on the LoCoMo benchmark, Kumiho achieves an F1 score of 0.565 and 93.3% judgment accuracy on LoCoMo-Plus, significantly outperforming baselines such as Gemini 2.5 Pro, while allowing seamless large language model substitution without pipeline modifications.
📝 Abstract
While individual components for AI agent memory exist in prior systems, their architectural synthesis and formal grounding remain underexplored. We present Kumiho, a graph-native cognitive memory architecture grounded in formal belief revision semantics. The structural primitives required for cognitive memory -- immutable revisions, mutable tag pointers, typed dependency edges, URI-based addressing -- are identical to those required for managing agent-produced work as versionable assets, enabling a unified graph-native architecture that serves both purposes. The central formal contribution is a correspondence between the AGM belief revision framework and the operational semantics of a property graph memory system, proving satisfaction of the basic AGM postulates (K*2--K*6) and Hansson's belief base postulates (Relevance, Core-Retainment). The architecture implements a dual-store model (Redis working memory, Neo4j long-term graph) with hybrid fulltext and vector retrieval. On LoCoMo (token-level F1), Kumiho achieves 0.565 overall F1 (n=1,986) including 97.5% adversarial refusal accuracy. On LoCoMo-Plus, a Level-2 cognitive memory benchmark testing implicit constraint recall, Kumiho achieves 93.3% judge accuracy (n=401); independent reproduction by the benchmark authors yielded results in the mid-80% range, still substantially outperforming all published baselines (best: Gemini 2.5 Pro, 45.7%). Three architectural innovations drive the results: prospective indexing (LLM-generated future-scenario implications indexed at write time), event extraction (structured causal events preserved in summaries), and client-side LLM reranking. The architecture is model-decoupled: switching the answer model from GPT-4o-mini (~88%) to GPT-4o (93.3%) improves end-to-end accuracy without pipeline changes, at a total evaluation cost of ~$14 for 401 entries.