Source-linked AI summary
MemOS: An Operating System for Memory-Augmented Generation (MAG) in Large Language Models
Zhiyu Li, Shichao Song, Hanyu Wang, Simin Niu, Ding Chen, Jiawei Yang, Chenyang Xi, Huayi Lai, Jihao Zhao, Yezhaohui Wang, Junpeng Ren, Zehao Lin, Jiahao Huo, Tianyi Chen, Kai Chen, Kehang Li, Zhiqiang Yin, Qingchen Yu, Bo Tang, Hongkang Yang, Zhi-Qin John Xu, Feiyu Xiong
TL;DR
Current LLMs lack unified, structured memory management, while existing approaches do not adequately support memory lifecycle evolution across heterogeneous forms. MemOS introduces a memory operating system with standardized abstractions and a memory-centric execution framework. The paper presents this architecture as groundwork for continual adaptation, long-term evolution, and cross-platform or multi-agent intelligence.
Problem
Current LLMs rely on difficult-to-manage parametric memory and context-limited activation memory, while existing memory systems lack unified lifecycle governance and fusion across roles or agents.
Method
MemOS unifies parametric, activation, and plaintext memory through lifecycle management, memory-centric execution, and the standardized MemCube abstraction.
Results
MemOS provides integrated scheduling, lifecycle management, structured storage, and transparent augmentation for heterogeneous memory types.
Takeaways & Limitations
MemOS lays groundwork for LLMs with continual learning, long-term evolution, integrated knowledge, and cross-task, cross-platform, multi-agent collaboration.
Takeaways & Limitations
Existing systems discussed in the paper still lack unified scheduling, lifecycle governance, and memory fusion across roles or agents.
Abstract
from arXiv · showhide
Large Language Models (LLMs) have emerged as foundational infrastructure in the pursuit of Artificial General Intelligence (AGI). Despite their remarkable capabilities in language perception and generation, current LLMs fundamentally lack a unified and structured architecture for handling memory. They primarily rely on parametric memory (knowledge encoded in model weights) and ephemeral activation memory (context-limited runtime states). While emerging methods like Retrieval-Augmented Generation (RAG) incorporate plaintext memory, they lack lifecycle management and multi-modal integration, limiting their capacity for long-term knowledge evolution. To address this, we introduce MemOS, a memory operating system designed for LLMs that, for the first time, elevates memory to a first-class operational resource. It builds unified mechanisms for representation, organization, and governance across three core memory types: parametric, activation, and plaintext. At its core is the MemCube, a standardized memory abstraction that enables tracking, fusion, and migration of heterogeneous memory, while offering structured, traceable access across tasks and contexts. MemOS establishes a memory-centric execution framework with strong controllability, adaptability, and evolvability. It fills a critical gap in current LLM infrastructure and lays the groundwork for continual adaptation, personalized intelligence, and cross-platform coordination in next-generation intelligent systems.
1 Introduction
Current LLMs lack robust, unified memory architecture: parametric memory is difficult to interpret, update, and transfer, while RAG provides only an ad hoc textual patch. MemOS addresses these limitations through lifecycle-managed, structured memory infrastructure.
- Current LLM architectures rely on implicit parametric memory embedded in model weights, making it difficult to interpret, update, or transfer.
- RAG incorporates external knowledge but lacks a structured, unified mechanism for memory management.
- MemOS treats memory as an operational resource with lifecycle management spanning generation, organization, utilization, and evolution.
- The proposed infrastructure provides structured representations, unified interfaces, version control, and access governance for evolving LLM memory.
2 Memory in Large Language Models
Research on LLM memory has progressed from classifying memory mechanisms to human-like persistence and systematic management. Existing systems still lack unified scheduling, lifecycle governance, and memory fusion across roles or agents.
- Memory Definition and Exploration: Memory research first classified mechanisms along parametric versus non-parametric and short-term versus long-term dimensions.
- Emergence of Human-like Memory: A second stage introduced systems emphasizing long-term persistence, context awareness, self-reflection, behavior continuity, and personalized memory modeling.
- Systematic Memory Management: A third stage added tool-based operations and OS-inspired governance, including explicit memory manipulation and paged context management.
- Systematic Memory Management: Existing systems still fall short of unified scheduling, lifecycle governance, and memory fusion across roles or agents.
3 MemOS Design Philosophy
MemOS proposes a memory-centric execution paradigm in which learning and inference become part of one unified process. It treats memory as a schedulable, governable resource that can be organized, evolved, migrated, and fused.
- Future LLM systems require structured memory modeling and unified management to support knowledge updates, state retention, and personalized adaptation.
- MemOS unifies learning and inference within a memory-driven process supporting structured storage, interface invocation, and lifecycle management.
- MemOS treats memory as a schedulable core resource and supports recomposition, migration, and fusion across agents, users, applications, and sessions.
- Access control, traceability, and interpretability are included as governance mechanisms for safe and compliant model operation.
4 MemOS
MemOS organizes parametric, activation, and plaintext memory within a unified operational framework. MemCube standardizes heterogeneous memory for scheduling, lifecycle management, governance, transformation, and reuse.
- 4.1 Types of Memory in MemOS: MemOS classifies memory into parametric, activation, and plaintext types with distinct representations, lifecycles, and invocation mechanisms.Together, they form a multi-layered structure for an intelligent agent’s cognitive system.
- 4.2 Memory Cube (MemCube) as a Core Resource: MemCube unifies heterogeneous memory through a standard abstraction supporting cross-type scheduling, lifecycle management, and structured fusion.It serves as the system’s fundamental encapsulation unit for memory resources with differing origins and invocation semantics.
- 4.2 Memory Cube (MemCube) as a Core Resource: MemCube combines semantic payloads with metadata describing memory identity, timestamps, origins, and semantic roles.Governance attributes add permissions, lifespan policies, priority levels, and compliance mechanisms such as sensitivity tags and access logging.
- 4.2 Memory Cube (MemCube) as a Core Resource: Behavioral indicators support adaptive transformations: frequently accessed plaintext can become activation templates, while stable knowledge can become parametric structures.Rarely used or outdated parameters can instead be externalized into editable plaintext.
- 4.3 MemOS Architecture: MemOS uses Interface, Operation, and Infrastructure layers to parse requests, schedule and evolve memory, and provide storage, governance, and interoperability.Its architecture manages the memory lifecycle through modular components unified by the MemoryCube abstraction.
- 4.4 System Execution Flow: The execution flow parses prompts into Memory API calls, selects relevant memory, injects it into context, and governs transformation, storage, and reuse.MemScheduler selects parametric, activation, or plaintext memory according to access patterns and scheduling policies.
5 Conclusion
MemOS introduces foundational infrastructure for managing heterogeneous memory in LLM applications. The paper outlines future work on cross-model sharing, self-evolving memory units, and decentralized memory exchange.
- 5 Conclusion: MemOS provides unified management for parametric, activation, and plaintext memory through MemCube, scheduling, lifecycle, storage, and augmentation modules.The authors associate these components with improved reasoning coherence, adaptability, and scalability in LLMs.
- 5 Conclusion: The authors envision modular memory resources and a decentralized memory marketplace supporting continual learning and long-term evolution.This vision centers next-generation AI systems on reusable and governable memory infrastructure.
- 5 Conclusion: Planned cross-LLM memory sharing would extend interoperability, reuse, compatibility, and trust mechanisms for transmitting memory across models and applications.The proposed Memory Interchange Protocol would define standard formats and compatibility rules.
- 5 Conclusion: Planned Self-Evolving MemBlocks would use usage feedback for self-optimization, reconstruction, and evolution, reducing manual maintenance and supervision.This direction targets memory units that adapt over time.
- 5 Conclusion: A scalable memory marketplace would support decentralized exchange, asset-level transactions, collaborative updates, and distributed evolution.The stated goal is a sustainable AI ecosystem centered on memory resources.