Conflict-Preserving Canonical Memory for Heterogeneous Agent Systems

Long-lived agents can retrieve memories that remain semantically relevant but are no longer valid. Existing memory systems largely optimize what to store and retrieve, leaving stale index entries and conflicting updates able to influence downstream actions. Volkov et al. introduced StateFuse, a canonical memory layer that separates retrieval from authority. This work extends it with applicability-aware conflict semantics, a resolution lifecycle, and live retrieval connectors. StateFuse records evidence, claims, retractions, conflicts, and resolutions in immutable canonical state, while external memory systems serve as disposable retrieval projections. Returned references are hydrated against current state before use, preventing a lagging or corrupted index from deciding what an agent should believe. Across four live memory backends, this boundary prevented every induced stale reference from being activated and improved safe-answer accuracy for two language models. On a natural-language subset of the official STALE benchmark, authoritative recovery improved task success under repository faults by 16.7 percentage points over raw Mem0; its margin over the strongest application-level defenses was inconclusive, with blind extraction and update linking remaining important bottlenecks. These results support post-retrieval validation as a practical safety boundary for heterogeneous agent memory, rather than as a replacement for semantic understanding.