Huawei Noah's Ark open-sources MindMemOS: an evolving memory operating layer for AI agents

Huawei's Noah's Ark Lab has open-sourced MindMemOS, a transferable and self-evolving memory operating layer designed for AI agents. The code is live on GitHub (github.com/mindscale-noah/MindMemOS) under the MIT License, and a cloud service on the project site mindmemos.cn is already open for registration, as reported by Chinese tech outlet QbitAI. The project targets a familiar pain point: every time a user switches agents or even sessions, the new agent behaves like a stranger, forcing users to re-explain preferences, project context, and hard-won lessons.

MindMemOS decouples memory from individual agents and models it with an entity-attribute-time three-dimensional structure that preserves both the latest state and the full evolution trajectory. Instead of storing memory as isolated text snippets or vector chunks, the system places "the user used to prefer X" and "the user now prefers Y" on the same timeline for the same entity and attribute, allowing semantic links between entities and traceable evolution of every attribute.

Two complementary memory-generation paths sit on top of this structure. MindVanilla needs no predefined modeling templates and ingests dialogue, text, and tool-execution traces, making it suitable for quickly absorbing open-domain information; MindSchema first defines which entities and attributes matter for a domain, then performs topic-segment splitting, attribute-level memory generation, equivalent-entity fusion, and graph merging. Retrieval is not a single vector-similarity lookup either: a Compact Search mechanism finds the shortest path in the memory graph that also covers the most information.

The system also introduces a Dreaming mechanism inspired by how humans consolidate memory. In an offline phase it identifies duplicates, conflicts, and evolutionary relationships, merges redundant content, archives facts superseded by newer ones, and records supersedes relations. On the FactConsolidation subset of MemoryAgentBench, Dreaming improved Single-hop and Multi-hop scores while moving about a fifth of active memories into archive — a case where less interference is itself a quality gain.

A Feedback mechanism routes user corrections back into the memory system. Users can point out an erroneous memory in natural language, or the system can detect implicit dissatisfaction, corrections, and new preferences in later conversations, then decide whether to add, update, archive, delete, or reinforce relevant memories. In the paper's PersonaMem-Evo case, Feedback extracted the genuinely stable preference from multiple interactions and added the right social-scenario boundary, changing downstream retrieval results.

Above memory, MindMemOS treats skills as a key application: it provides cloud-based skill registration, version chains, sync, and rollback, and collects real task trajectories through a unified interface. Once trajectories accumulate, the system evolves skills unsupervised from repeated success and failure patterns (MindEvolve-Unsup) or with task scores that reinforce effective behavior and suppress common errors (MindEvolve-Sup). On SpreadsheetBench-Verified, a benchmark of 400 real spreadsheet tasks, an unoptimized initial skill actually underperformed using no skill at all, while unsupervised evolution lifted success from 51.3% to 55.3% and adding task scores reached 57.2%.

MindMemOS also reports leading results on two mainstream long-term memory benchmarks: on LoCoMo, MindMemOS-Modeling posted an Overall Accuracy of 94.03, the highest among all compared methods, and on the long-term personalization benchmark PersonaMem, the MindSchema configuration scored 70.63% while MindVanilla reached 67.74%. Architecturally, the system decouples the agent-access layer, memory-algorithm layer, and memory-structure layer, with the open-source code offering a FastAPI HTTP interface, a Python SDK, a CLI, Skills, and an OpenClaw plugin.

The direct payoff of this decoupling is portability: memory no longer lives inside a single agent's private implementation but can be maintained as a long-term asset owned by a user, project, or organization and reused across applications and agent frameworks. MindMemOS has also been integrated into the automatic algorithm design platform LLM4AD_Next, converting feedback from algorithm search into storable, retrievable, reusable experience assets with Auto and Manual memory scheduling modes.

The bigger signal is that the competition is shifting from how well an agent performs a single task to what it retains after a hundred tasks. As more agent frameworks plug into memory operating layers like MindMemOS, memory is positioned to become independent infrastructure reused across applications, and versioned, rollback-capable skill management may become standard for long-lived agents. The next thing to watch is cloud adoption and how the memory and skill evolution mechanisms hold up inside other agent ecosystems.