MemOS: A Memory-Centric Operating System for Evolving and Adaptive Large Language Models

As Large Language Models (LLMs) become essential in the pursuit of Artificial General Intelligence (AGI), they face significant challenges related to memory management. Current LLMs primarily rely on fixed knowledge encoded in their weights and ephemeral context during operation, which complicates the retention and updating of information over time. While techniques like Retrieval-Augmented Generation (RAG) attempt to incorporate external knowledge, they often lack structured memory management. This results in issues such as forgetting past interactions, limited adaptability, and fragmented memory across different platforms. Ultimately, existing LLMs do not treat memory as a manageable, persistent, or shareable resource, which restricts their practical applications.

To overcome these memory limitations, researchers from MemTensor (Shanghai) Technology Co., Ltd., Shanghai Jiao Tong University, Renmin University of China, and the Research Institute of China Telecom have developed MemOS. This memory operating system positions memory as a primary resource within language models. Central to MemOS is MemCube, a unified memory abstraction that manages parametric, activation, and plaintext memory. MemOS facilitates structured, traceable, and cross-task memory handling, enabling models to adapt continuously, internalize user preferences, and maintain behavioral consistency. This paradigm shift transforms LLMs from passive generators into dynamic systems capable of long-term learning and cross-platform coordination.

As AI systems become increasingly complex—managing multiple tasks, roles, and data types—language models must evolve beyond mere text comprehension to include memory retention and continuous learning. Current LLMs’ lack of structured memory management hinders their adaptability and growth. MemOS addresses this by treating memory as a core, schedulable resource, allowing for long-term learning through structured storage, version control, and unified memory access. Unlike traditional training methods, MemOS supports a continuous “memory training” approach that merges learning with inference. It also emphasizes governance, ensuring traceability, access control, and safe utilization in evolving AI systems.

MemOS redefines memory in language models, categorizing it into three distinct types:

• Parametric Memory: Knowledge embedded in model weights through pretraining or fine-tuning.
• Activation Memory: Temporary internal states, such as key-value caches and attention patterns used during inference.
• Plaintext Memory: Editable, retrievable external data, including documents or prompts.

These memory types interact within a unified framework known as MemCube, which encapsulates both content and metadata, allowing for dynamic scheduling, versioning, access control, and transformation across memory types. This structured system empowers LLMs to adapt, recall pertinent information, and evolve their capabilities, transforming them into more than just static generators.

At the heart of MemOS is a three-layer architecture:

• Interface Layer: Handles user inputs and parses them into memory-related tasks.
• Operation Layer: Manages the scheduling, organization, and evolution of different memory types.
• Infrastructure Layer: Ensures safe storage, access governance, and cross-agent collaboration.

All interactions within the system are mediated through MemCubes, enabling traceable, policy-driven memory operations. Through modules like MemScheduler, MemLifecycle, and MemGovernance, MemOS maintains a continuous and adaptive memory loop—from the moment a user sends a prompt, to memory injection during reasoning, to storing useful data for future use. This design enhances the model’s responsiveness and personalization while ensuring that memory remains structured, secure, and reusable.

In conclusion, MemOS is a memory operating system that positions memory as a central, manageable component in LLMs. Unlike traditional models that rely predominantly on static weights and short-term states, MemOS introduces a unified framework for managing parametric, activation, and plaintext memory. At its core is MemCube, a standardized memory unit that supports structured storage, lifecycle management, and task-aware memory augmentation. This system enables more coherent reasoning, adaptability, and cross-agent collaboration. Future objectives include facilitating memory sharing across models, developing self-evolving memory blocks, and establishing a decentralized memory marketplace to support continual learning and intelligent evolution.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter.