跳转至

文章背景与核心概要

在人工智能领域，AI智能体（AI Agents）极大地依赖记忆系统来维持上下文、实现用户个性化以及保持长期的适应性。然而，传统的记忆框架在开发完成后通常保持静态，缺乏通过持续交互来有机优化其记忆结构、组织策略或程序性技能的能力。

为了突破这一局限，本文作者推出了 MindMemOS，这是一个可移植且具备自进化能力的记忆操作系统层，旨在通过统一的“实体-属性-时间”结构来管理开放世界的信息。该系统具备场景自适应架构、MindMemEvolve 演化算法、“做梦”（Dreaming）机制、隐式人机协同反馈以及 MindSkillEvolve 技能演化算法等多项核心创新。在基准测试中，MindMemOS 在 LOCOMO 上取得了 94.03% 的准确率，在 PersonaMem 上达到了 70.63%，同时将 SpreadsheetBench 的成功率较初始基线提升了 9.2 个百分点。

MindMemOS: A Portable and Self-Evolving Memory Operating Layer for AI Agents

arXiv: 2608.12428 [cs.AI]
Submitted: August 12, 2026
Authors: Kaichao Liang, Yuqi Cui, Hao Kong, Xinyuan Huang, Guohaotian Hou, Qingcan Kang, Liang Chen, Yiyang Yin, Ke Ye, Jiaquan Guo, Da Chen, Lingan Zeng, Yixing Peng, Rong Yao, Shixiong Kai, Mingxuan Yuan

📌 Summary

AI agents rely heavily on memory to maintain context, user personalization, and long-term adaptability. However, traditional memory frameworks remain static post-development, lacking the capacity to organically refine their memory structures, organization policies, or procedural skills through ongoing interaction.

To overcome this limitation, the authors introduce MindMemOS, a portable and self-evolving memory operating layer designed to manage open-world information via a unified entity-property-time structure. Key innovations and performance metrics include: * Scenario-Adaptive Architecture: Dynamically shapes memory modeling, discovers higher-order patterns, and refines stored knowledge autonomously. * MindMemEvolve Algorithm: Utilizes validation-driven evolutionary search to optimize target memory schemas. * Dreaming Mechanism: Consolidates historical memories by resolving conflicts and merging redundant entries. * Implicit Human-in-the-Loop Feedback: Corrects potentially inaccurate or misaligned memories automatically. * MindSkillEvolve Algorithm: Transforms execution trajectories into progressively refined, reusable agent skills. * State-of-the-Art Performance: Achieves 94.03% accuracy on the LOCOMO benchmark and 70.63% on PersonaMem, while boosting SpreadsheetBench success rates by 9.2 percentage points over initial baselines.

👥 Authors & Affiliations

Kaichao Liang, Yuqi Cui, Hao Kong, Xinyuan Huang, Guohaotian Hou, Qingcan Kang, Liang Chen, Yiyang Yin, Ke Ye, Jiaquan Guo, Da Chen, Lingan Zeng, Yixing Peng, Rong Yao, Shixiong Kai, Mingxuan Yuan

📄 Abstract

记忆是 AI 智能体维持经验积累、保持个性化以及在长期交互中实现自适应的核心组件。然而，现有的记忆系统在开发后往往固定不变，这限制了它们通过持续使用来调整记忆模型、组织策略和程序性知识的能力。

我们提出了 MindMemOS，这是一个可移植且具备自进化能力的记忆操作系统层，它利用统一的“实体-属性-时间”（entity property timestructure）结构来组织开放世界的信息。MindMemOS 支持场景自适应的记忆建模、高阶模式发现、自主记忆优化以及持续的技能演化。

其 MindMemEvolve 算法采用验证驱动的演化搜索来针对目标场景优化记忆模式，而“做梦”机制则通过合并冗余记录和解决冲突来整合积累的记忆。此外，隐式的纠错反馈充当了人机协同（human-in-the-loop）信号，用于识别和修正潜在的不准确或不对齐的记忆。其 MindSkillEvolve 算法进一步将智能体的执行轨迹转化为可重用且逐步精炼的技能。

MindMemOS 在 LOCOMO 上取得了 94.03% 的准确率，在 PersonaMem 上达到了 70.63%。与初始技能基线相比，MindSkillEvolve 将 SpreadsheetBench 的成功率提升了 9.2 个百分点。

Memory is a core component of AI agents, enabling them to accumulate experience, maintain personalization, and adapt over long-term interactions. However, existing memory systems often remain fixed after development, limiting their ability to adapt their memory models, organization strategies, and procedural knowledge through continued use.

We present MindMemOS, a portable and self-evolving memory operating layer that organizes open-world information using a unified entity property timestructure. MindMemOS supports scenario-adaptive memory modeling, higher-order pattern discovery, autonomous memory refinement, and continuous skill evolution.

Its MindMemEvolve algorithm employs validation-driven evolutionary search to optimize memory schemas for target scenarios, while dreaming consolidates accumulated memories by merging redundant records and resolving conflicts. In addition, implicit corrective feedback serves as a human-in-the-loop signal for identifying and revising potentially inaccurate or misaligned memories. Its MindSkillEvolve algorithm further transforms agent execution trajectories into reusable and progressively refined skills.

MindMemOS achieves 94.03% accuracy on LOCOMO and 70.63% on PersonaMem. MindSkillEvolve improves SpreadsheetBench success by 9.2 percentage points over the initial-skill baseline.

更多资源与链接： * 全文访问： 查看 PDF | HTML（实验性） | TeX 源码 * 许可证： 知识共享署名 4.0 license icon * 引用与指标： 谷歌学术 | Semantic Scholar | NASA ADS