Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents
Just-in-Time Memory curates raw agent trajectories at read time and beats write-time memory on three benchmarks.
Most agent memory systems distill a finished trajectory into a fixed artifact before future queries are known, discarding detail and creating long-horizon credit assignment. Just-in-Time Memory keeps raw trajectories and, once the current task is known, synthesizes a compact task-adaptive payload that can be trained from immediate success. On ALFWorld, WebShop, and tau2-bench it outperforms no-memory, heuristic, and learned write-time methods by 16.2, 16.3, and 3.9 absolute success points over the strongest baseline. An untrained curator is already competitive, and training the curator adds further gains.
- Keeps raw trajectories and curates memory at read time
- Trains the curator from immediate task success
- Gains 16.2, 16.3, and 3.9 points on three benchmarks
- An untrained curator is already competitive with write-time baselines
From the original article
1 min read at huggingface.co
Agentic memory systems reuse past experience to improve future performance, yet most existing designs curate memory at write time: once a task is completed, its trajectory is distilled into a fixed artifact, such as a reflection, workflow, skill, or reasoning strategy, that is later retrieved by similarity.
By Yefan Zhou, Yang Li, Zeyu Leo Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.