记忆是重构而非检索：面向LLM智能体的图记忆
英文摘要
Researchers propose MRAgent, a framework that improves long-horizon memory reasoning for LLM agents. It uses a Cue-Tag-Content graph representation and an active reconstruction mechanism to dynamically retrieve and prune memory paths during inference. This moves beyond the static retrieve-then-reason paradigm, adapting memory access to intermediate reasoning evidence. Experiments show up to 23% performance improvement over baselines and reduced token and runtime costs. The work demonstrates efficient memory reconstruction for complex agent tasks.
中文摘要
研究者提出了MRAgent框架，通过关联记忆图与主动重构机制提升LLM智能体的长程记忆推理。它采用Cue-Tag-Content图表示和动态记忆路径探索与剪枝，突破静态检索-推理的局限，使记忆访问能适应推理过程中的中间证据。实验显示相比强基线性能提升最高23%，并降低了token和运行时间消耗，验证了在高复杂性任务中记忆重构的有效性。
关键要点
MRAgent introduces a Cue-Tag-Content graph memory representation for LLM agents.
MRAgent为LLM智能体引入Cue-Tag-Content图记忆表示。
Active reconstruction mechanism dynamically prunes retrieval paths based on intermediate reasoning evidence.
主动重构机制根据推理中间证据动态剪枝检索路径。
Achieves up to 23% performance gain on benchmark tasks and reduces token and runtime costs.
在基准任务上性能提升最高23%，同时降低token和运行时间成本。