JitMem Curates LLM Agent Memory at Read Time, Not Write Time

"Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time," opens a new paper on Hugging Face by Yefan Zhou, Yang Li, Zeyu Leo Liu, Semih Yavuz, and Shafiq Joty, before arguing for the opposite move.

The authors' complaint about write-time curation is blunt. Compressing a finished trajectory into a reflection, workflow, or skill forces the system "to decide what is worth remembering before the future query is known, irreversibly discarding information and producing a query-independent summary that must serve many possible downstream tasks." Their alternative, called Just-in-Time Memory or JitMem, keeps raw traces and only synthesizes a compact payload once a new task shows up.

On three agent benchmarks, JitMem "consistently outperforms no-memory agents as well as heuristic and learned write-time memory methods, improving over the strongest baseline by 16.2, 16.3, and 3.9 absolute success-rate points" on ALFWorld, WebShop, and τ²-bench respectively. The result that will draw the most attention is buried at the end of the abstract: "even an untrained curator is already competitive with or surpasses these baselines, showing that task-adaptive read-time curation itself is a major source of the gain; training the curator further compounds the improvement."

The paper landed on the same day as a burst of adjacent agent-memory work on our tracker. No inference-cost, storage, or latency figures accompany the accuracy claims in the abstract, so the operational trade-off of running a curator on every read is left for the full text.