MindMemOS - Huawei Noah's Ark open-source AI Agent memory operating system - AiBoss

What is MindMemOS?

MindMemOS is an open-source AI Agent memory operating system from Huawei Noah's Ark Labs, addressing the pain point of agents forgetting what they've used. The system decouples memory from individual agents into independent assets that can be reused across frameworks, using a three-dimensional entity-attribute-time structure to store the latest state and complete evolution trajectory. MindMemOS achieves co-evolution of memory and the system through offline memory organization via Dreaming, user correction signal mining via Feedback, and continuous skill evolution based on real execution trajectories.

Main functions of MindMemOS

3D memory structureOrganize memory using three-dimensional coordinates of entities, attributes, and time, and track the evolution of attributes and semantic relationships between entities.
Dual-mode memory retrievalMindVanilla enables fast template-free import of open domains; MindSchema provides pre-defined entity and attribute rules to adapt to different business scenarios.
Compact SearchThe outer Agentic module plans the retrieval path, while the inner module performs multi-path searches such as entity reverse lookup, multi-hop relationship expansion, and timeline tracing.
Dreaming Offline OrganizationThe offline phase identifies duplicates, conflicts, and evolutionary relationships, merges redundant versions, archives invalid versions, and supplements higher-order patterns.
Feedback loopSupports explicit and implicit feedback, and performs add, update, archive, delete or reinforce operations after determining the signal type.
Skill EvolutionBased on the actual execution trajectory, it automatically extracts successful strategies and failure modes, generates skill modification plans, and iterates through versions.

How to use MindMemOS

Clone the repository and prepare the configuration.：cp .env.example .env and cp config/mindmemos/dev.example.yaml config/mindmemos/dev.yamlConfigure chat_model_router, embed_model_router, rerank_model_router, and API key in the configuration file.
Start service:implement make dev Start the full Docker dependency stack and FastAPI service (default port 8000), or make dev-core Only enable core dependencies.
Install and configure the SDK：pip install mindmemos-sdk,run mindmemos auth Configure base_url, api_key, and user_id.
Adding and retrieving memoriesCall via SDK client.memory.add() Write dialogue or facts, using client.memory.search() Perform the search.
Submit feedback: call client.memory.feedback() or CLI mindmemos memory feedback Submit explicit or implicit correction signals.
Trigger Dreaming:implement mindmemos memory dreaming Alternatively, you can use the dreaming API for offline memory consolidation.
Registration and Evolution Skill:pass mindmemos skill register Register a local Skill and use it. mindmemos skill evolve Version evolution is triggered by execution trajectory.

MindMemOS project address

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

MindMemOS's core advantages

Memory can be transferredDecouple the Agent access layer, memory algorithm layer, and memory structure layer, and reuse memory as an independent long-term asset across applications.
Structured Recall LeadingLoCoMo achieved an overall accuracy of 94.03% (SOTA) and PersonaMem achieved 70.63%, significantly outperforming solutions such as Mem0, MemU, Zep, and EverOS.
Less is more in memory qualityDreaming improves question-and-answer accuracy by up to 10.3 percentage points while compressing 19.4%-23.5% of active memory.
Skill Self-evolutionThe initial skill, without optimization, was actually lower than the baseline without skill, while MindEvolve-Sup increased the task success rate from 51.3% to 57.2%, proving the value of turning experience into rules.
Feedback-driven evolutionThe Feedback mechanism uses user correction signals to influence memory retrieval and search strategies.
Multiple access methodsSupports local deployment (Docker + FastAPI), cloud API, Python SDK, CLI, and OpenClaw plugin.

Comparison of MindMemOS with similar competing products

Dimension	MindMemOS	Mem0
Memory structure	A 3D graph structure of entity-attribute-time, tracing the evolution trajectory.	Text fragments/vector blocks, flat storage
Memory retrieval	MindVanilla + MindSchema dual-mode, with preset domain rules.	Fixed template extraction is highly versatile but lacks flexibility.
Memory organization	Dreaming features offline active consolidation, automatically resolving conflicts and redundancy.	There is no offline processing mechanism; it relies on model judgment during querying.
Feedback loop	Explicit + Implicit Feedback: Reverse Optimization of Extraction and Retrieval Strategies	Supports explicit feedback, primarily modifying individual memory entries.
Skill Evolution	Skill version based on real trajectory automatic evolution	No Skill evolution capability
Cross-Agent Migration	Decoupled architecture, supporting multiple frameworks to share the same memory library.	The memory is bound to the Mem0 platform.
LoCoMo accuracy	94.03%	64.20%
Open source license	MIT	MIT

Application Scenarios of MindMemOS

Personalized AI AssistantIt retains user preferences, project background, and past pitfalls across sessions, avoiding the need to re-train agents each time.
Multi-Agent Collaboration PlatformThe same memory can be shared and reused by different agents such as OpenClaw, Hermes, Claude Code, and OpenHands.
Automatic Algorithm Design (LLM4AD)The effective experiences and pitfall avoidance guidelines from algorithm search are systematized into a three-layer memory that is searchable and reusable.
Complex office automationBased on Skill Evolution, we continuously optimize the execution strategies for repetitive tasks such as spreadsheet processing and document editing.
Long-term user profile buildingTrack changes in user attributes over time to support personalized recommendations and social planning.