一篇综述（参见图1）从三个视角——形式、功能和动态——整理了AI智能体中“记忆”这一核心能力。“记忆”一词往往因人而异，含义混乱，因此这种梳理非常有用。《AI智能体时代的记忆》https://arxiv.org/abs/2512.13564

AI智能体时代的记忆

来源：https://arxiv.org/abs/2512.13564 作者：Yuyang Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+Y), Shichun Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+S), Yanwei Yue (https://arxiv.org/search/cs?searchtype=author&query=Yue,+Y), Guibin Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+G), Boyang Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+B), Fangyi Zhu (https://arxiv.org/search/cs?searchtype=author&query=Zhu,+F), Jiahang Lin (https://arxiv.org/search/cs?searchtype=author&query=Lin,+J), Honglin Guo (https://arxiv.org/search/cs?searchtype=author&query=Guo,+H), Shihan Dou (https://arxiv.org/search/cs?searchtype=author&query=Dou,+S), Zhiheng Xi (https://arxiv.org/search/cs?searchtype=author&query=Xi,+Z), Senjie Jin (https://arxiv.org/search/cs?searchtype=author&query=Jin,+S), Jiejun Tan (https://arxiv.org/search/cs?searchtype=author&query=Tan,+J), Yanbin Yin (https://arxiv.org/search/cs?searchtype=author&query=Yin,+Y), Jiongnan Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+J), Zeyu Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+Z), Zhongxiang Sun (https://arxiv.org/search/cs?searchtype=author&query=Sun,+Z), Yutao Zhu (https://arxiv.org/search/cs?searchtype=author&query=Zhu,+Y), Hao Sun (https://arxiv.org/search/cs?searchtype=author&query=Sun,+H), Boci Peng (https://arxiv.org/search/cs?searchtype=author&query=Peng,+B), Zhenrong Cheng (https://arxiv.org/search/cs?searchtype=author&query=Cheng,+Z), Xuanbo Fan (https://arxiv.org/search/cs?searchtype=author&query=Fan,+X), Jiaxin Guo (https://arxiv.org/search/cs?searchtype=author&query=Guo,+J), Xinlei Yu (https://arxiv.org/search/cs?searchtype=author&query=Yu,+X), Zhenhong Zhou (https://arxiv.org/search/cs?searchtype=author&query=Zhou,+Z), Zewen Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+Z), Jiahao Huo (https://arxiv.org/search/cs?searchtype=author&query=Huo,+J), Junhao Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+J), Yuwei Niu (https://arxiv.org/search/cs?searchtype=author&query=Niu,+Y), Yu Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+Y), Zhenfei Yin (https://arxiv.org/search/cs?searchtype=author&query=Yin,+Z), Xiaobin Hu (https://arxiv.org/search/cs?searchtype=author&query=Hu,+X), Yue Liao (https://arxiv.org/search/cs?searchtype=author&query=Liao,+Y), Qiankun Li (https://arxiv.org/search/cs?searchtype=author&query=Li,+Q), Kun Wang (https://arxiv.org/search/cs?searchtype=author&query=Wang,+K), Wangchunshu Zhou (https://arxiv.org/search/cs?searchtype=author&query=Zhou,+W), Yixin Liu (https://arxiv.org/search/cs?searchtype=author&query=Liu,+Y), Dawei Cheng (https://arxiv.org/search/cs?searchtype=author&query=Cheng,+D), Qi Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+Q), Tao Gui (https://arxiv.org/search/cs?searchtype=author&query=Gui,+T), Shirui Pan (https://arxiv.org/search/cs?searchtype=author&query=Pan,+S), Yan Zhang (https://arxiv.org/search/cs?searchtype=author&query=Zhang,+Y), Philip Torr (https://arxiv.org/search/cs?searchtype=author&query=Torr,+P), Zhicheng Dou (https://arxiv.org/search/cs?searchtype=author&query=Dou,+Z), Ji-Rong Wen (https://arxiv.org/search/cs?searchtype=author&query=Wen,+J), Xuanjing Huang (https://arxiv.org/search/cs?searchtype=author&query=Huang,+X), Yu-Gang Jiang (https://arxiv.org/search/cs?searchtype=author&query=Jiang,+Y), Shuicheng Yan (https://arxiv.org/search/cs?searchtype=author&query=Yan,+S)

查看 PDF (https://arxiv.org/pdf/2512.13564)

摘要：记忆已经成为，并将继续是，基于基础模型的智能体的核心能力。随着智能体记忆研究的快速扩展并吸引前所未有的关注，该领域也变得日益碎片化。现有属于智能体记忆范畴的工作在动机、实现和评估协议上往往存在显著差异，而松散定义的记忆术语的激化进一步模糊了概念清晰性。传统的分类法（如长期/短期记忆）已被证明不足以捕捉当代智能体记忆系统的多样性。本文旨在提供当前智能体记忆研究的最新图景。我们首先明确界定了智能体记忆的范围，并将其与相关概念（如LLM记忆、检索增强生成（RAG）和上下文工程）区分开来。然后，我们通过形式、功能和动态这三个统一视角审视智能体记忆。从形式的角度，我们识别了智能体记忆的三种主要实现方式，即token级记忆、参数化记忆和潜在记忆。从功能的角度，我们提出了一种更细粒度的分类法，区分了事实记忆、经验记忆和工作记忆。从动态的角度，我们分析了记忆如何随时间形成、演化和检索。为支持实际开发，我们编制了记忆基准测试和开源框架的全面摘要。除了整合现有研究，我们还展望了新兴研究前沿，包括记忆自动化、强化学习集成、多模态记忆、多智能体记忆和可信性问题。我们希望本综述不仅作为现有工作的参考，而且作为重新思考记忆作为未来智能体设计中的一等原语的概念基础。

提交历史

来自：Guibin Zhang 查看邮件 [v1] 2025年12月15日星期一 17:22:34 UTC (28,501 KB) [v2] 2026年1月13日星期二 09:33:57 UTC (28,509 KB)