Researchers at the National University of Singapore have developed MRAgent, a framework that abandons the static 'retrieve-then-reason' approach in favor of a dynamic memory reconstruction process. This multi-step process is integrated into the reasoning process of large language models, significantly reducing token consumption and runtime costs.
MRAgent treats memory as an interactive environment, allowing the agent to explore multiple candidate retrieval paths and iteratively optimize its search. The framework uses a 'Cue-Tag-Content' mechanism to organize its database, enabling efficient two-stage retrieval and reducing noise in the context window.
In tests on the LoCoMo and LongMemEval industry benchmarks, MRAgent outperformed other frameworks, including A-MEM and LangMem, with a significant reduction in prompt token consumption and runtime costs.
