New agentic memory framework uses 118K tokens per query. LangMem burns through 3.26M.

Researchers at the National University of Singapore developed MRAgent, a new agentic memory framework that dynamically develops memory based on accumulating evidence. MRAgent significantly reduces token consumption and runtime costs compared to other approaches by implementing active memory reconstruction and organizing its database using a Cue-Tag-Content mechanism. Tested on industry benchmarks, MRAgent outperformed other frameworks in resolving queries on long-horizon tasks. Its efficient on-demand behavior and autonomous context evaluation contribute to its effectiveness in practice. The framework's automated distillation pipeline simplifies the preparation of the underlying memory database for developers.