research(memory): Kumiho formal belief revision for graph memory — AGM-compliant versioned edge updates (arXiv:2603.17244)

[bug-ops]

Paper

Title: Graph-Native Cognitive Memory for AI Agents: Formal Belief Revision Semantics for Versioned Memory Architectures
arXiv: https://arxiv.org/abs/2603.17244
Published: 2026-03-18

Key Technique

Kumiho proves formal AGM belief revision postulates (K2–K6 + Hansson's Relevance + Core-Retainment) for a property graph memory system. Implements immutable revision edges + mutable tag pointers for versioned belief update. Dual-store: Redis working memory + Neo4j long-term graph. Hybrid fulltext + vector retrieval.

Benchmark: LoCoMo F1 = 0.447 (SOTA across retrieval categories) + 97.5% adversarial refusal accuracy. 100% AGM compliance across 49 scenarios × 7 postulates.

Why Relevant to Zeph

Directly addresses the open limitation in MAGMA temporal edge versioning (Partial status): Zeph's current graph edges never expire because LLM extraction produces varying relation strings (works_at vs works_on), so conflict detection by (source,target,relation) never fires. Kumiho's belief revision semantics provide the formal framework for correct invalidation: when a new fact contradicts an old one, the old edge gets a valid_to timestamp regardless of exact relation string match.

Implementation sketch for MAGMA:

• Add that applies AGM K*6 (Vacuity) — new facts about same entity pair in same domain get validity-checked against prior edges
• Replace exact (source,target,relation) match with semantic similarity + entity-pair matching
• Tag edges with revision chains: pointer for audit trail
• Reference Kumiho's formal postulates as implementation contract

Priority Rationale

P2: fixes a known architectural gap (temporal edge versioning marked Partial since v0.15.1) with a principled, formally verified approach. Zero speculative features — fills documented limitation.