StateFuse: Deterministic Conflict-Preserving Memory for Multi-Agent Systems
06:00 · July 8, 2026 · arXiv cs.AI RSS
Agent systems accumulate conflicting observations across branches, retries, and replicas, yet many practical memory layers still collapse disagreement behind overwrite rules that are difficult to inspect or correct. We present StateFuse, a conflict-aware replicated memory contract built on standard OpSet/CRDT merge. StateFuse does not introduce a new join algebra; it defines an agent-facing semantics layer with immutable history, explicit conflict objects, exact and semantic correction handles (claim_id / claim_ref), deterministic predicate contracts, and projection-time resolution that cannot rewrite replicated state. We evaluate StateFuse against flat multi-value, raw-log, provenance-style, and collapsed baselines under matched resolver and verification policies. On a 282-question official conflict-bearing MemoryAgentBench slice, the compared methods tie on answer accuracy, but conflict-preserving surfaces keep contradictions visible while collapsed surfaces do not. In a controlled agent loop with uniform verification, preserving ambiguity enables safer abstention and correction than early collapse. A correction-handle ablation further shows that semantic handles matter when exact prior identifiers are unavailable. The resulting claim is narrow: StateFuse is best supported as a safer public memory contract for contradiction surfacing, abstention, and auditable correction, not as a universal accuracy gain.
Summary
StateFuse addresses a recurring problem in multi-agent systems: branches, retries, and replicas frequently produce incompatible observations about the same entity, yet conventional memory layers often resolve these through opaque overwrite rules that hide the disagreement from downstream agents. The system supplies a contract layer on top of standard OpSet/CRDT merge semantics rather than a new convergence algebra. It stores an immutable history of claims, retractions, and decisions, surfaces explicit conflict objects at projection time, and supplies two correction mechanisms—exact claim identifiers for local edits and semantic claim references for cross-replica targets—while restricting resolvers to choosing among visible candidates or abstaining.
The contract further includes deterministic predicate rules for normalization and claim-reference derivation, together with a strict separation between replicated state and any task-specific projection. This design keeps contradictions durable and inspectable instead of collapsing them at merge time. When evaluated on a 282-question conflict-bearing slice of MemoryAgentBench, StateFuse and other conflict-preserving surfaces match the answer accuracy of collapsed baselines, yet only the former expose contradictions for every task. In controlled agent loops that apply identical verification budgets, surfaces that preserve ambiguity support safer abstention and subsequent correction, whereas early collapse leaves the agent with materially worse operating points.
Ablation results indicate that semantic correction handles become important precisely when exact prior identifiers are unavailable. The authors therefore position StateFuse narrowly as a safer public memory contract for surfacing contradictions, enabling abstention, and supporting auditable edits, rather than as a general route to higher accuracy.
Why it matters
This research is highly relevant for Dutch AI practitioners developing multi-agent systems, as it directly addresses the need for transparent and auditable AI memory architectures. By preserving data conflicts rather than overwriting them, StateFuse aligns strongly with EU and Dutch priorities for ethical, explainable, and safe AI deployments.
