arXiv cs.CL· Zhiqian Xia, Jielong Tang, Ling-An Zeng, Shixiong Kai, Rong Yao, Hao Kong, Xiao Dong, Wei Wang, Mingxuan Yuan, Siyu Xia, Haifeng Xia·· 13 小时前AI 评分32
Uni-Mem：用统一在线策略优化学习长程智能体记忆管理器
Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization
AI 导读
Uni-Mem 提出端到端在线强化学习框架，让记忆构建与检索在共享策略下联合学习，并配套 Stage-Decoupled GRPO 交替优化两阶段、Event-Wise Credit Assignment（ECA）将反馈定位到具体事件。在 LoCoMo 和 LongMemEval 上，Uni-Mem 优于最强的基于 RL 的记忆方法。
正文 · 原文
Authors:Zhiqian Xia, Jielong Tang, Ling-An Zeng, Shixiong Kai, Rong Yao, Hao Kong, Xiao Dong, Wei Wang, Mingxuan Yuan, Siyu Xia, Haifeng Xia
Abstract:Long-term agent memory is shifting from predefined workflows toward learnable policies for memory construction and retrieval. Existing learning-based methods either optimize construction alone or combine retrieval-side online reinforcement learning (RL) with offline construction supervision, limiting co-adaptation between the two stages. In this paper, we explore end-to-end online RL for a unified memory manager (Uni-Mem) that learns both capabilities under a shared policy. However, joint optimization with existing RL algorithms such as GRPO faces confounded group comparisons and sparse feedback over long memory-management trajectories. To this end, we propose Stage-Decoupled GRPO, which alternates construction and retrieval optimization while controlling variation from the other stage within each rollout group, yielding stage-specific learning signals. We further introduce Event-Wise Credit Assignment (ECA), which localizes feedback to responsible events and propagates event-level advantages to their constituent operations. Experiments on LoCoMo and LongMemEval show that Uni-Mem outperforms the strongest RL-based memory, demonstrating the effectiveness of jointly learning memory construction and retrieval via Stage-Decoupled GRPO and ECA.
| Comments: | 21 pages, 8 figures |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.29671 [cs.CL] |
| (or arXiv:2610.29671v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.29671 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Jielong Tang [view email]
[v1]
Wed, 21 Oct 2026 12:48:10 UTC (961 KB)
来源：arXiv cs.CL · arxiv.org