JitMem shifts agent memory to read time
Salesforce AI Research’s JitMem stores successful agent trajectories in raw form, then curates task-specific memory only when a new task arrives. The paper reports gains over write-time memory methods on ALFWorld, WebShop, and τ²-bench (https://arxiv.org/abs/2609.27334).
JitMem makes a compelling case that agent memory should preserve experiences until their future use is known, though read-time curation shifts complexity and cost into every retrieval.
- –Task-conditioned summaries avoid irreversible information loss from fixed reflections or skills.
- –JitMem improved over the strongest baseline by 16.2 points on ALFWorld, 16.3 on WebShop, and 3.9 on τ²-bench.
- –The curator trains against immediate task success, simplifying credit assignment compared with write-time memory.
- –Compact payloads reduced input tokens by 50.3%–56.3% and executor steps by 28.4%–31.4% in the paper’s tests.
- –Production systems will still need strong retrieval, storage controls, and safeguards against retaining sensitive raw trajectories.
DISCOVERED
1h ago
2026-09-26
PUBLISHED
1h ago
2026-09-26
RELEVANCE
AUTHOR
Discover AI