MemOS: A Memory OS for AI System

Large Language Models (LLMs) have become an essential infrastructure forArtificial General Intelligence (AGI), yet their lack of well-defined memorymanagement systems hinders the development of long-context reasoning, continualpersonalization, and knowledge consistency.Existing models mainly rely onstatic parameters and short-lived contextual states, limiting their ability totrack user preferences or update knowledge over extended periods.WhileRetrieval-Augmented Generation (RAG) introduces external knowledge in plaintext, it remains a stateless workaround without lifecycle control orintegration with persistent representations.Recent work has modeled thetraining and inference cost of LLMs from a memory hierarchy perspective,showing that introducing an explicit memory layer between parameter memory andexternal retrieval can substantially reduce these costs by externalizingspecific knowledge. Beyond computational efficiency, LLMs face broaderchallenges arising from how information is distributed over time and context,requiring systems capable of managing heterogeneous knowledge spanningdifferent temporal scales and sources. To address this challenge, we proposeMemOS, a memory operating system that treats memory as a manageable systemresource. It unifies the representation, scheduling, and evolution ofplaintext, activation-based, and parameter-level memories, enablingcost-efficient storage and retrieval. As the basic unit, a MemCube encapsulatesboth memory content and metadata such as provenance and versioning. MemCubescan be composed, migrated, and fused over time, enabling flexible transitionsbetween memory types and bridging retrieval with parameter-based learning.MemOS establishes a memory-centric system framework that bringscontrollability, plasticity, and evolvability to LLMs, laying the foundationfor continual learning and personalized modeling.