# MemOS: A Memory OS for AI System

**Authors:** Zhiyu Li, Shichao Song, Chenyang Xi, Hanyu Wang, Chen Tang, Simin Niu, Ding Chen, Jiawei Yang, Chunyu Li, Qingchen Yu, Jihao Zhao, Yezhaohui Wang, Peng Liu, Zehao Lin, Pengyuan Wang, Jiahao Huo, Tianyi Chen, Kai Chen, Kehang Li, Zhen Tao  
**arXiv:** [2507.03724](https://arxiv.org/abs/2507.03724)  
**Date:** 2025-07-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.03724v1/

## Key points

- Large Language Models (LLMs) are critical for Artificial General Intelligence (AGI) and excel in natural language processing tasks
- LLMs have evolved to handle structured code generation, cross-modal reasoning, multi-turn dialogue, and complex planning
- Efficient memory management systems are crucial due to the increasing size and complexity of models
- MemOS proposes a memory operating system that treats memory as a manageable resource through MemCubes
- LLMs are expected to become persistent agents embedded in workflows, accumulating interaction histories and adapting over time
- The evolution of memory systems in LLMs is categorized based on object type, form, temporal aspects, and retention duration
- MemOS establishes a memory-centric framework for controllability, plasticity, and evolvability in LLMs towards achieving AGI capabilities

## Summary

Large Language Models (LLMs) have emerged as a critical component of Artificial General Intelligence (AGI), showcasing near-human performance in various natural language processing tasks. With the advancement of Transformer architecture and self-supervised pretraining, LLMs have expanded their capabilities to structured code generation, cross-modal reasoning, multi-turn dialogue, and complex planning. As models continue to grow in size and complexity, they are positioned as a key pathway towards AGI. In response to the evolving landscape of LLMs, the need for efficient memory management systems has become increasingly apparent. Current models lack well-defined memory structures, hindering long-context reasoning, continual personalization, and knowledge consistency. While existing approaches like Retrieval-Augmented Generation (RAG) introduce external knowledge in plain text, they lack lifecycle control and integration with persistent representations. To address these challenges, MemOS proposes a memory operating system that treats memory as a manageable resource. By unifying plaintext, activation-based, and parameter-level memories within MemCubes - encapsulating both content and metadata - MemOS enables cost-efficient storage and retrieval. This framework allows for flexible transitions between different memory types and bridges retrieval with parameter-based learning. Looking ahead, the presence of LLMs is expected to expand temporally and spatially. Models will transition from stateless tools to persistent agents embedded in long-running workflows, accumulating interaction histories and adapting internal states over time. Spatially, LLMs are becoming foundational intelligence layers across users, platforms, and ecosystems - necessitating efficient organization,
storage,
and retrieval of knowledge. The evolution of memory systems in large language models is highlighted through systematic classifications based on parameters such as object type (personal vs. system), form (parametric vs. non-parametric), temporal aspects (short-term vs. long-term), retention duration distinguishing sensory memory from short-term to long-term memory. Overall, MemOS establishes a memory-centric framework that brings controllability, plasticity, and evolvability to LLMs - laying the foundation for continual learning and personalized modeling towards achieving AGI capabilities.