MindMemOS – Huawei Noah's Ark Lab's Open-Source AI Agent Memory Operating System Operational

MindMemOS is an open-source AI agent memory operating system from Huawei Noah's Ark Lab that addresses the problem of agents forgetting everything after use. It decouples memory from individual agents into an independent asset that can be reused across frameworks, using a three-dimensional entity–attribute–time structure to preserve the latest state and the complete evolution history. MindMemOS enables the joint evolution of memory and systems through Dreaming for offline memory organization, Feedback for mining user correction signals, and Skill Evolution for continuously evolving skills based on real execution trajectories.

Follow us on WeChat and reply “开源” to join the AI open-source project community
3D memory structure: Organize memory using three-dimensional entity, attribute, and time coordinates while tracking attribute evolution and semantic relationships between entities.
Dual-mode memory extraction: MindVanilla enables fast, template-free ingestion for open-domain data; MindSchema prespecifies entity and attribute rules for different business scenarios.
Compact Search retrieval: The outer Agentic module plans retrieval paths, while the inner layer performs multi-path searches such as reverse entity lookup, multi-hop relationship expansion, and timeline tracking.
Dreaming offline organization: During offline processing, identify duplicate, conflicting, and evolutionary relationships, merge redundancies, archive obsolete versions, and supplement higher-level patterns.
Feedback loop: Support explicit and implicit feedback, determine the signal type, and then perform add, update, archive, delete, or reinforcement operations.
Skill Evolution: Automatically extract successful strategies and failure patterns from real execution trajectories, generate Skill modification plans, and iterate versions.

Clone the repository and prepare the configuration: cp .env.example .env and cp config/mindmemos/dev.example.yaml config/mindmemos/dev.yaml. Set chat_model_router, embed_model_router, rerank_model_router, and the API key in the configuration file.
Start the services: Run make dev to start the complete Docker dependency stack and FastAPI service (default port 8000), or make dev-core to start only the core dependencies.
Install and configure the SDK: Run pip install mindmemos-sdk, then run mindmemos auth to configure base_url, api_key, and user_id.
Add and retrieve memories: Use client.memory.add() through the SDK to write conversations or facts, and client.memory.search() to perform retrieval.
Submit feedback: Call client.memory.feedback() or run mindmemos memory feedback in the CLI to submit explicit or implicit correction signals.
Trigger Dreaming: Run mindmemos memory dreaming or call the dreaming API for offline memory consolidation.
Register and evolve Skills: Register a local Skill with mindmemos skill register, then use mindmemos skill evolve to trigger version evolution based on execution trajectories.

Official website: https://mindmemos.cn/
GitHub repository: https://github.com/mindscale-noah/MindMemOS

Portable memory: Decouple the agent integration layer, memory algorithm layer, and memory structure layer, allowing memory to serve as an independent long-term asset that can be reused across applications.
Leading structured recall: Overall LoCoMo accuracy reaches 94.03% (SOTA), while PersonaMem reaches 70.63%, significantly outperforming solutions such as Mem0, MemU, Zep, and EverOS.
Higher memory quality through less: Dreaming compresses 19.4%-23.5% of active memories while improving question-answering accuracy by up to 10.3 percentage points.
Self-evolving Skills: Unoptimized initial Skills perform below the no-Skill baseline, while MindEvolve-Sup raises task success rates from 51.3% to 57.2%, demonstrating the value of turning experience into rules.
Feedback-driven evolution: The Feedback mechanism feeds user correction signals back into memory extraction and retrieval strategies.
Multiple integration methods: Supports local deployment (Docker + FastAPI), cloud APIs, a Python SDK, CLI, and the OpenClaw plugin.

Personalized AI assistants: Retain user preferences, project context, and past lessons across sessions, avoiding the need to retrain the agent from scratch each time.
Multi-agent collaboration platforms: Allow different agents, including OpenClaw, Hermes, Claude Code, and OpenHands, to share and reuse the same memory store.
Automated algorithm design (LLM4AD): Systematize effective experience and failure-avoidance guidance from algorithm searches into three layers of searchable, reusable memory.
Complex office automation: Continuously optimize execution strategies for repetitive tasks such as spreadsheet processing and document editing through Skill Evolution.
Long-term user profiling: Track how user attributes change over time to support personalized recommendations and social planning.