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

Researchers at the National University of Singapore developed MRAgent, an agentic memory framework that dynamically reconstructs memory during reasoning instead of relying on static retrieve-then-reason pipelines. The system reportedly uses about 118K tokens per query, far below LangMem’s 3.26M, reducing both context noise and runtime costs.

MRAgent reflects a broader shift from passive retrieval toward active memory systems that adapt as an agent reasons through complex tasks. If the approach generalizes, it could make long-running AI agents cheaper and more reliable by reducing irrelevant context and enabling more targeted recall.