New LLM Agent Memory System Improves Long-Context Reasoning

Large Language Model (LLM) agents often struggle with long-context reasoning due to the accumulation of redundant or irrelevant information in their memory. A new framework, Strategic Forgetting for Agent Memory Systems (SF-AMS), addresses this by introducing a dynamic, utility-driven mechanism for managing memory. Instead of static retrieval or simple decay, SF-AMS assesses the long-term importance of memory units based on usage redundancy and temporal signals. This approach creates a hierarchical memory structure that prioritizes stable, entity-consistent information while actively filtering out noise. Additionally, it integrates semantic and entity-level signals through Composite Importance Scoring to improve retrieval robustness. Experiments demonstrate that SF-AMS consistently outperforms existing state-of-the-art baselines, showing substantial improvements in multi-hop reasoning, temporal reasoning, and open-domain tasks across various LLM backbones like Qwen2.5-7B and GPT-4o-mini. This highlights the critical role of dynamic memory importance modeling for reliable long-context reasoning in LLM agents.