不要询问LLM追踪新鲜度：一种确定性的内存冲突解决策略
Reliable Post-Retrieval Assembly for Agent Memory: Separating Evidence Extraction from Policy Execution
AI总结：
针对基于LLM的内存系统中事实冲突解决性能低下的问题，提出用候选提取加Python max(serial)的确定性聚合替代LLM判断，在单跳任务上提升10.8个百分点，并扩展到多跳任务。
摘要：
基于LLM的内存系统越来越多地维护随时间演变的事实，其中一个反复出现的失败是冲突解决：当一个事实有多个矛盾的值时，智能体应该返回哪个？MemoryAgentBench (MAB; Hu et al., 2026) 在其FactConsolidation任务中明确了这一点：事实被编号，反事实具有更高的序号，并且智能体被告知较新的事实具有较大的序号。然而，每个已发布的系统表现不佳：HippoRAG-v2在单跳（FC-SH）上达到54%，BM25 48%，Mem0 18%，而时间知识图谱Zep/Graphiti仅为7%。多跳几乎未解决（22个系统中最多7%）。我们认为瓶颈在于组装步骤：基线将冲突解决留给LLM介导的检索或生成，而不是版本感知的聚合。一个匹配设置的比较（相同的主干、检索、分块、TOP_K）表明，用候选提取加Python max(serial)替换LLM判断答案流水线，在FC-SH上（gpt-4o-mini）获得+10.8分的提升，从6K时的+8分扩大到262K时的+21分。这是一个全流水线效应（解析器、提示、格式和温度共同变化）；隔离解析器是未来的工作。该配方在FC-SH上达到78.0%（gpt-4o-mini）、94.8%（gpt-4o），在FC-MH上达到30.2%（gpt-4o-mini，使用gpt-4o时升至51.5%），通过每跳确定性的Self-Ask扩展。在匹配的262K下，它比HippoRAG-v2高出+28分，比已发布的最佳FC-MH结果高出+20分。这一含义对该子领域具有纠正作用：冲突解决的瓶颈是组装（检索后聚合），而不是存储。一个LongMemEval知识更新检查表明，该机制从max(serial)移植到max(timestamp)，但仅与LLM判断持平（57.8% vs 64.4%，n=45）：确定性聚合是当前值冲突的正确原语，并且必须与问题类型感知处理组合，以实现更广泛的内存问答。
英文摘要：
LLM-based memory systems can retrieve relevant evidence yet still fail when answer generation entangles semantic filtering, conflict resolution, prior suppression, and output generation in one step. We study this failure as a problem of post-retrieval assembly. In the MemoryAgentBench (MAB) release used here, FactConsolidation explicitly states that newer facts have larger serial numbers, yet the best reported retrieval/memory result is 54% single-hop and all 22 reported systems score at most 7% multi-hop. We evaluate a structured assembly interface in which an LLM first extracts semantically matching evidence into a candidate representation and a separate stage executes the required answer policy. At 262K, this pipeline reaches 82%/93% single-hop and 27%/41% multi-hop with gpt-4o-mini/gpt-4o, exceeding every result reported in the MAB v3 FactConsolidation comparison. This is a task-level result, not a claim that the evaluated memory architectures are broadly inferior. A controlled whole-pipeline comparison, with identical backbone, retrieved top-10 evidence, chunking, and n=100 per cell, improves single-hop accuracy by 10.8 percentage points (pp) on average and 21 pp at 262K. A targeted comparison using the same extraction setup shows that changing only the final policy executor contributes 2.0 pp on average and 0 pp at 262K. Most of the gain therefore comes from separating evidence identification from final policy execution rather than from the freshness operator itself. A LongMemEval check finds no significant overall advantage (26/45 versus 29/45; paired exact McNemar p=0.45), bounding the result to current-value questions with explicit version metadata. The evidence identifies post-retrieval assembly as a distinct reliability boundary between retrieval and answer generation.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
