Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents
The paper’s bet is that agents should decide what memory matters only after the next task is known.
JitMem keeps raw past trajectories instead of compressing them at write time.
When a new task arrives, it curates a compact memory payload from retrieved traces for that task.
The authors report gains over the strongest baseline of 16.2 points on ALFWorld, 16.3 on WebShop, and 3.9 on τ²-bench.
They also find that even an untrained read-time curator is competitive with or better than write-time memory methods.
HF Daily Papers' note
JitMem keeps raw past trajectories instead of compressing them at write time.
When a new task arrives, it curates a compact memory payload from retrieved traces for that task.
The authors report gains over the strongest baseline of 16.2 points on ALFWorld, 16.3 on WebShop, and 3.9 on τ²-bench.
They also find that even an untrained read-time curator is competitive with or better than write-time memory methods.
HF Daily Papers' note
score 5