StateFuse: Deterministic Conflict-Preserving Memory for Multi-Agent Systems

🕒 Published (v1): 2026-07-07 05:06 UTC · Source: Arxiv · link

Why this paper was selected

Deterministic conflict-preserving memory for multi-agent systems; fixes silent overwrite in distributed agents

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

StateFuse is a conflict-aware memory contract for multi-agent systems built on standard OpSet/CRDT merge that preserves contradictions as explicit, public objects rather than collapsing them via silent overwrite. It introduces exact and semantic correction handles, deterministic predicate contracts, and projection-bounded resolution. The core empirical claim is narrow: StateFuse provides a safer public memory contract for contradiction surfacing, abstention, and auditable correction—not a universal accuracy improvement.

Problem

Multi-agent systems operating across branches, retries, and replicas accumulate conflicting observations, but most practical memory layers silently overwrite disagreements via opaque overwrite rules. This causes hidden branch disagreement, stale observations surviving validated corrections, and unsafe actions taken from collapsed memory surfaces. Existing agent memory evaluations (RAG, MemGPT, ReAct-style) focus on retrieval quality and long-horizon behavior, not on what happens when concurrent beliefs contradict each other or when a later correction must invalidate a stale claim without erasing it.

Method

StateFuse defines a five-component agent-facing semantics layer on top of a standard OpSet merge substrate (merge is plain set union: \(\text{merge}(O_1, O_2) = O_1 \cup O_2\)):

1. Immutable replicated history over four operation types: EvidenceAdded, ClaimAdded, ClaimRetracted, DecisionAdded. Claims are keyed by (namespace, subject, predicate).
2. Dual correction handles: a claim_id (exact per-assertion identifier) and a claim_ref (deterministic semantic handle derived from the claim key and predicate-governed value contract), enabling cross-replica correction even when the original writer's opaque ID is unavailable.
3. Deterministic predicate contracts: a registry per predicate specifying functional/multi-valued nature, deterministic normalize(value) and equal(left, right) functions, and contract-checker validation before deployment.
4. Explicit ConflictSet objects emitted at projection/materialization time for any functional predicate with multiple distinct active values; retracted claims never contribute candidates.
5. Bounded resolver authority: resolvers operate only at projection time and may choose a candidate, abstain, or fail closed—they cannot mutate base memory. An optional LLM resolver is restricted to structured JSON I/O.

Materialization is \(O(|O|)\) in the op-set size; conflict detection is linear in active claims; projection is linear in active keys.

Key Contributions

• A conflict-aware memory contract with immutable history, explicit public conflict surfacing, exact and semantic correction handles, deterministic predicate contracts, and projection-bounded resolution.
• A matched-information evaluation design where strong flat multi-value and raw-log baselines receive the same resolver family and downstream verification budget, preventing confounds from surface-dependent tool allocation.
• Evaluation on a 282-question official MemoryAgentBench conflict-bearing slice, a uniform-verification synthetic agent loop, and a correction-handle ablation separating claim_id from claim_ref behavior.

Results

• Official MemoryAgentBench (282 conflict-bearing questions): StateFuse (conservative), StateFuse core, flat multi-value, provenance surface, raw-log, and collapsed-latest-write all reach 97.5% final accuracy in aggressive mode—no accuracy advantage for StateFuse.
• Under conservative abstention, all conflict-preserving surfaces (StateFuse, flat MV, provenance) converge to 64.9% accuracy, 100% conflict recall, 2.1% false certainty; raw-log and collapsed surfaces expose 0% conflict recall.
• Uniform-verification agent loop (50 tasks): all conservative non-collapsing surfaces reach 100% post-verification success, 0% false-confident actions; collapsed surface reaches only 60% post-verification success with 40% false-confident actions.
• Correction-handle ablation (13 tasks each): claim_ref achieves 100% recoverability and 100% no-resurrection on semantic-target tasks; claim_id drops to 76.9% accuracy and 76.9% no-resurrection on the same tasks, and achieves 0% recoverability on semantic targets where the exact ID is unavailable.

Limitations

• No accuracy improvement over strong flat multi-value baselines on the official benchmark; the tie means the method cannot claim general accuracy superiority.
• The uniform-verification agent loop is synthetic and controlled; real-world multi-agent workloads may differ substantially.
• The official MemoryAgentBench conflict-bearing slice contains no correction-target or gold-abstention cases, so recoverability and abstention quality metrics are N/A there.
• Bounded compaction and authenticated sync are treated as secondary engineering mechanisms, not evaluated as independent contributions.
• False-certainty and abstention metrics are explicitly not "calibration" in the formal probabilistic sense.
• Scale experiments are absent; complexity analysis is theoretical.

Relevance to Agentic AI / LLM Agents

StateFuse directly addresses a gap in multi-agent architectures where memory shared across branching, retrying, or replicated agents needs a principled conflict-management contract rather than ad-hoc overwrite. The key insight—that preserving contradictions as first-class public objects enables safer abstention and auditable correction—is directly applicable to agent orchestration frameworks that spawn parallel subagents whose beliefs may diverge. The semantic claim_ref handle solves a concrete cross-replica correction problem that arises in distributed LLM agent pipelines where opaque per-session identifiers are lost. The matched-information evaluation protocol also sets a methodological precedent for future agent memory benchmarks to disentangle information preservation, conflict structuring, and downstream policy effects.