Large language models are increasingly expected to answer queries grounded in long-horizon histories, such as multi-session dialogues and lengthy documents. To support that, they turn to external memory. A new preprint on arXiv, "All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents" (arXiv:2610.21147), targets the limitations of current memory methods.
The abstract notes that existing memory methods often optimize retrieval and other components, but it is cut off before detailing the proposed solution. The title itself points to the contribution: a process-rewarded reinforcement learning framework for memory agents. The paper appears to position this as an "all-in-one" alternative to piecemeal memory optimization.
Given the truncated abstract, this digest can only report the problem statement and the proposed method name. Full evaluation of the approach would require reading the complete paper.