MemOS: A Memory OS for AI System

TL;DR

MemOS introduces a memory operating system for LLMs that treats memory as a first-class resource, unifying plaintext, activation, and parameter memories under the MemCube abstraction. It provides a three-layer architecture (Interface, Operation, Infrastructure) with modules for memory reading, scheduling, lifecycle management, and governance, enabling controllable, evolvable memory across tasks and platforms. Through comprehensive evaluations on long-context memory, personalization, and memory acceleration, MemOS demonstrates superior performance and robustness against baselines, highlighting the practical impact of memory-centric reasoning and continual learning. The work envisions a memory-driven AI ecosystem with modular memory assets, cross-LLM sharing, and decentralized memory marketplaces to support scalable, evolving intelligent agents.

Abstract

Large Language Models (LLMs) have become an essential infrastructure for Artificial General Intelligence (AGI), yet their lack of well-defined memory management systems hinders the development of long-context reasoning, continual personalization, and knowledge consistency.Existing models mainly rely on static parameters and short-lived contextual states, limiting their ability to track user preferences or update knowledge over extended periods.While Retrieval-Augmented Generation (RAG) introduces external knowledge in plain text, it remains a stateless workaround without lifecycle control or integration with persistent representations.Recent work has modeled the training and inference cost of LLMs from a memory hierarchy perspective, showing that introducing an explicit memory layer between parameter memory and external retrieval can substantially reduce these costs by externalizing specific knowledge. Beyond computational efficiency, LLMs face broader challenges arising from how information is distributed over time and context, requiring systems capable of managing heterogeneous knowledge spanning different temporal scales and sources. To address this challenge, we propose MemOS, a memory operating system that treats memory as a manageable system resource. It unifies the representation, scheduling, and evolution of plaintext, activation-based, and parameter-level memories, enabling cost-efficient storage and retrieval. As the basic unit, a MemCube encapsulates both memory content and metadata such as provenance and versioning. MemCubes can be composed, migrated, and fused over time, enabling flexible transitions between memory types and bridging retrieval with parameter-based learning. MemOS establishes a memory-centric system framework that brings controllability, plasticity, and evolvability to LLMs, laying the foundation for continual learning and personalized modeling.

Figures (10)

• Figure 1: MemOS achieves state-of-the-art performance across all benchmarks. This figure provides a comprehensive summary of our evaluation results, illustrating the personalized response rate on PreFEval(0 turns and 10 turns), the precision score on PersonaMem, the overall mean score on LongMemEval, and the overall mean LLM judge score on the LoCoMo benchmark. The MemOS (MemOS-1031) consistently ranks first in all categories, significantly outperforming strong baselines such as MIRIX, Mem0, Zep, Memobase, MemU, and Supermemory. Full detailed metric breakdowns are provided in Table \ref{['tab:new_locomo']}, Table \ref{['tab:perfeval']}, Table \ref{['tab:longmemeval']} and Table \ref{['tab:personamem']}.
• Figure 2: Categorization of LLM knowledge, including the memory hierarchy. The explicit memories, extracted from model activations, lie half-way between raw data and model parameters, so a dotted line is used to indicate that they may or may not be regarded as parameters. Reproduced from memory3_Yang_2024.
• Figure 3: Illustration of the evolution of memory systems in large language models, highlighting the progression from definition and exploration, to human-like memory development, and to tool-based memory management.
• Figure 4: Phased transitions in model performance: from pretraining and post-training to the Mem-training stage. MemOS serves as the foundational infrastructure enabling the next era of scaling laws.
• Figure 5: Transformation paths among three types of memory, forming a unified, controllable, and evolvable memory space.