Managing long-context dependencies remains a primary bottleneck in LLM agents, as redundant and irrelevant information can degrade multi-step reasoning. The Strategic Forgetting for Agent Memory Systems (SF-AMS) framework is proposed to maintain compact, high-utility memory by modeling the long-term importance of memory units.
SF-AMS replaces static retrieval and heuristic decay with a utility-driven survival mechanism that updates memory importance based on usage redundancy and temporal signals, inducing a hierarchical memory structure that prioritizes stable entity-consistent information while filtering noise.
Moreover, Composite Importance Scoring integrates semantic and entity-level signals to improve retrieval robustness. Experiments on LoCoMo and LongMemEval-s show consistent gains over strong state-of-the-art baselines, including LightMem, MemO, and A-Mem.
The largest improvement appears in multi-hop reasoning under Qwen2.5-7B, where SF-AMS achieves a 9.65 F1 increase over the strongest baseline, followed by temporal reasoning under GPT-4o-mini with a 6.91 F1 increase and open-domain tasks with a 6.53 F1 increase, demonstrating strong cross-backbone generalization. These results indicate that modeling memory importance as a dynamic utility signal is critical for reliable long-context reasoning.
Blogger's Review: The SF-AMS framework presents a groundbreaking approach to dynamically manage the importance of memory in LLM agents, paving the way for efficient reasoning. Future research could further explore optimizing this mechanism for more complex applications.