JitMem: Just-in-Time Memory for LLM Agents

"Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents." An agent-memory method from Salesforce AI Research (Yefan Zhou and Yang Li co-first, with Zeyu Leo Liu, Semih Yavuz and Shafiq Joty) that moves curation from write time to read time. Most agent-memory systems distill each finished task into a fixed artifact (a reflection, insight, workflow, skill or reasoning strategy) and retrieve it later by similarity, so they decide what to keep before the next query is known. Learning such a write-time curator also suffers from delayed credit, because a storage decision pays off only when some later task retrieves it. JitMem stores raw trajectories instead. For each new task a retriever pulls past trajectories from the bank and a curator model condenses them into a compact, task-specific payload for a frozen executor; trajectories the executor judges successful go back into the bank. Because the payload is used on the same task, the curator's reward is that task's outcome, and the paper trains a Qwen3-8B curator with GRPO for 100 steps without the task grouping that learned write-time curators such as SkillOS need. BAAI's JAM, posted five days later, also keeps raw history and assembles memory at query time; JitMem keeps the curator separate from the executor, so one trained curator serves several executors.

The evaluation covers ALFWorld, WebShop and τ²-bench with Qwen3-8B, Gemini-2.5-Pro and GPT-5.4 as frozen executors, against a no-memory agent and three write-time methods (ReasoningBank, MemP, SkillOS). With RL-trained Qwen3-8B curators on both sides, JitMem beats SkillOS by 77.4 to 61.2 success rate on ALFWorld and 32.8 to 16.5 on WebShop; with Gemini-2.5-Pro executing, the margins are 86.2 to 80.2 and 50.5 to 41.3. Read-time curation helps even untrained: with Qwen3-8B as executor and curator, untrained JitMem reaches 60.5 on ALFWorld against 55.7 for ReasoningBank. τ²-bench has no training split, so only training-free variants were compared there: with GPT-5.4 as executor and curator, JitMem scores 75.6 against 71.7 for ReasoningBank, with the largest per-domain gain on Telecom (+11.0) and no memory method beating the no-memory agent beyond variance on Airline and Retail. Paired with a GPT-5.4 executor, a curator trained with Qwen3-8B as its executor comes within 1.4 points on ALFWorld of one trained directly with GPT-5.4. The compact payloads cut executor input tokens by 50.3–56.3% and steps by 28.4–31.4% relative to write-time methods. No code was linked at filing.