Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures

Dimensions overall score 8.0

• Kumiho achieves 0.565 overall F1 score on the LoCoMo benchmark (n=1,986) including 97.5% adversarial refusal accuracy.result 95%
  Evidencepartial

  On LoCoMo (token-level F1), Kumiho achieves 0.565 overall F1 (n=1,986) including 97.5% adversarial refusal accuracy.

  Implicationpartial

  Directly stated in abstract with specific numeric results

  Verificationpartial

  partial

• Kumiho achieves 93.3% judge accuracy on LoCoMo-Plus benchmark testing implicit constraint recall (n=401).result 95%
  Evidencepartial

  On LoCoMo-Plus, a Level-2 cognitive memory benchmark testing implicit constraint recall, Kumiho achieves 93.3% judge accuracy (n=401)

  Implicationpartial

  Directly stated in abstract with specific numeric results

  Verificationpartial

  partial

• Kumiho substantially outperforms all published baselines on LoCoMo-Plus, with the best baseline (Gemini 2.5 Pro) achieving only 45.7% accuracy.result 90%
  Evidencepartial

  still substantially outperforming all published baselines (best: Gemini 2.5 Pro, 45.7%)

  Implicationpartial

  Direct comparison with specific baseline results provided

  Verificationpartial

  partial

• The architecture implements a dual-store model with Redis working memory and Neo4j long-term graph storage.technical 90%
  Evidencepartial

  The architecture implements a dual-store model (Redis working memory, Neo4j long-term graph)

  Implicationpartial

  Directly stated architectural implementation details

  Verificationpartial

  partial

• Kumiho's formal contribution includes proving satisfaction of the basic AGM postulates (K*2-K*6) and Hansson's belief base postulates.technical 90%
  Evidencepartial

  proving satisfaction of the basic AGM postulates (K*2--K*6) and Hansson's belief base postulates (Relevance, Core-Retainment)

  Implicationpartial

  Direct statement of formal contribution with specific postulates named

  Verificationpartial

  partial

• Three architectural innovations drive Kumiho's results: prospective indexing, event extraction, and client-side LLM reranking.method 85%
  Evidencepartial

  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.

  Implicationpartial

  Directly stated as innovations driving results, though causal relationship requires some inference

  Verificationpartial

  partial

• The architecture is model-decoupled, allowing switching from GPT-4o-mini to GPT-4o without pipeline changes, improving accuracy from ~88% to 93.3%.technical 85%
  Evidencepartial

  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

  Implicationpartial

  Direct statement of model-decoupling with specific accuracy improvements

  Verificationpartial

  partial

• The structural primitives required for cognitive memory are identical to those required for managing agent-produced work as versionable assets.technical 80%
  Evidencepartial

  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

  Implicationpartial

  Direct statement about architectural unification, though the equivalence claim requires some interpretation

  Verificationpartial

  partial

Competitive landscape

Kumiho is a graph-native cognitive memory architecture that enhances AI agents' memory capabilities through formal belief revision semantics.