SF-AMS: Strategic Forgetting for Structured Memory in LLM Agent

SF-AMS：大语言模型智能体中用于结构化记忆的策略性遗忘机制

Managing long-context dependencies remains a primary bottleneck in LLM agents, as redundant and irrelevant information can degrade multi-step reasoning. 管理长上下文依赖关系仍然是大语言模型（LLM）智能体的主要瓶颈，因为冗余和无关信息会降低多步推理的质量。

Strategic Forgetting for Agent Memory Systems (SF-AMS) is proposed as a framework for maintaining compact high-utility memory by modeling the long-term importance of memory units. 本文提出了“智能体记忆系统的策略性遗忘”（SF-AMS）框架，通过对记忆单元的长期重要性进行建模，旨在维护紧凑且高价值的记忆。

SF-AMS replaces static retrieval and heuristic decay with a utility-driven survival mechanism that updates memory importance from usage redundancy and temporal signals, inducing a hierarchical memory structure that prioritizes stable entity-consistent information while filtering noise. SF-AMS 取代了传统的静态检索和启发式衰减方法，采用了一种由效用驱动的生存机制。该机制根据使用冗余度和时间信号更新记忆重要性，从而构建出一种分层记忆结构，在过滤噪声的同时优先保留稳定的实体一致性信息。

On top of this, Composite Importance Scoring integrates semantic and entity level signals to improve retrieval robustness. 在此基础上，复合重要性评分（Composite Importance Scoring）整合了语义和实体层面的信号，进一步提高了检索的鲁棒性。

Experiments on LoCoMo and LongMemEval-s show consistent gains over strong state of the art baselines including LightMem, MemO, and A-Mem. 在 LoCoMo 和 LongMemEval-s 上的实验表明，该方法在 LightMem、MemO 和 A-Mem 等强基准模型的基础上取得了持续的性能提升。

The largest improvement appears in multi-hop reasoning under Qwen2.5-7B where SF-AMS achieves plus 9.65 F1 over the strongest baseline, followed by temporal reasoning under GPT-4o-mini plus 6.91 F1, and open-domain tasks plus 6.53 F1, demonstrating strong cross-backbone generalization. 性能提升最显著的是 Qwen2.5-7B 模型下的多跳推理任务，SF-AMS 较最强基准提升了 9.65 F1；其次是 GPT-4o-mini 下的时间推理任务（提升 6.91 F1）和开放域任务（提升 6.53 F1），这证明了该方法具有强大的跨模型骨干泛化能力。

These results show that modeling memory importance as a dynamic utility signal is critical for reliable long-context reasoning. 这些结果表明，将记忆重要性建模为动态效用信号，对于实现可靠的长上下文推理至关重要。