O-Mem: Omni Memory System for Personalized, Long Horizon, Self-Evolving Agents

WANG PIAOHONG, Motong Tian, Jiaxian Li, Yuan Liang, Yuqing Wang, Qianben Chen, Tiannan Wang, Zhicong Lu, Jiawei Ma, Yuchen Eleanor Jiang, Wangchunshu Zhou

Keywords: Memory Framework, Agent

Abstract: Recent advancements in LLM-powered agents have demonstrated significant potential in generating human-like responses; however, they continue to face challenges in maintaining long-term interactions within complex environments, primarily due to limitations in contextual consistency and dynamic personalization. Existing memory systems often depend on semantic grouping and the retrieval of past interaction groupings, which can overlook semantically irrelevant yet critical user information and introduce retrieval noise. To address these issues, we propose O-Mem, a novel memory framework based on active user profiling that dynamically extracts and updates user characteristics and event records from interactions. O-Mem supports hierarchical retrieval of persona attributes and topic-related context, enabling more adaptive and coherent personalized responses. Additionally, we introduce a new dataset designed to evaluate personalized long-text generation in memory-augmented agents. Experiments across three personalized tasks demonstrate that O-Mem consistently improves long-term human–AI interaction by scaling memory-time within interactions.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 5511