MemOS is an open-source operating system designed to equip Large Language Models (LLMs) with long-term memory capabilities, enabling more context-aware and personalized interactions. It targets LLM developers and researchers seeking to enhance model performance through structured memory management, offering significant improvements in reasoning tasks, particularly temporal reasoning.

How It Works

MemOS employs a modular architecture called MemCube, which supports various memory types: textual, activation (KVCache), and parametric (e.g., LoRA weights). This allows for flexible integration and management of different memory stores. A higher-level orchestration layer, MOS (Memory Operating System), manages multiple MemCubes and provides a unified API for memory operations, including storage, retrieval, and user-specific memory management. This approach aims to provide a structured and extensible framework for LLM memory augmentation.

Quick Start & Requirements

• Install: pip install MemoryOS
• Optional Dependencies: Install with feature groups like MemoryOS[tree-mem], MemoryOS[mem-reader], MemoryOS[mem-scheduler].
• External Dependencies: Ollama CLI for Ollama support. PyTorch (CUDA recommended) for transformer-based functionalities.
• Examples: Download example code and data with memos download_examples.

Highlighted Details

• Achieves 38.98% average score improvement over OpenAI baseline on LOCOMO benchmark.
• Temporal reasoning accuracy improved by 159% compared to OpenAI baseline.
• Supports Textual Memory, Activation Memory (KVCache), and Parametric Memory (LoRA weights).
• Features Memory-Augmented Generation (MAG) for unified memory operations.

Maintenance & Community

• Active development with recent releases and paper publications in 2025.
• Community support available via GitHub Issues, Discussions, Pull Requests, Discord, and WeChat.
• Links: GitHub, Discord, Website

Licensing & Compatibility

• Licensed under the Apache 2.0 License.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The project is currently in a "Preview" release (v1.0 Stellar), indicating potential for ongoing changes and stabilization. Specific performance claims are based on benchmark results detailed in their publications.