正在加载页面…
浏览 ArkGraph，选择本次要执行的步骤。
This paper introduces Just-In-Time Agent Memory (JAM), which constructs query-specific context by exploring complete stored histories at runtime. A fixed Memorizer organizes raw sessions into a hierarchical workspace with navigational summaries, while a trainable Researcher searches, opens directories, and browses evidence. Memory-Gym synthesizes evidence-grounded tasks across nine task types and six domains. Researcher training combines verified-trajectory supervised fine-tuning with hint-guided reinforcement learning. Evaluations compare JAM with retrieval, ahead-of-time memory systems, and trained memory agents on conversational memory, long-document reasoning, and multi-hop tasks. Additional analyses examine code-domain transfer, component ablations, runtime scaling, human data-quality audits, inference variability, and efficiency.
Preserving full histories and searching them for each question can recover details lost by advance summarization. The reported results suggest stronger answers across several memory tasks and transfer to code comprehension, with lower latency than MemAgent but higher query costs than one-shot systems. The paper also records imperfect synthetic-data quality, longer exploration for HotpotQA, and no consistent benefit from larger Memorizer models.
论文中的结论可在「研究结论」中查看。