Huawei Noah's Ark Lab Open-Sources MindMemOS: A Transferable, Self-Evolving Memory Operating Layer for AI Agents

Huawei Noah's Ark Lab introduced MindMemOS, an open-source memory operating layer for AI agents that addresses the problem of agent memory being non-transferable and non-evolving. Unlike traditional memory systems that store memories as text or vector chunks, MindMemOS uses an entity-property-time structure to capture both the latest state and the full evolution of facts. It offers two memory generation paths: MindVanilla for open-domain information and MindSchema for domain-specific modeling with customizable extraction rules. The system includes a Compact Search mechanism that goes beyond simple vector similarity to plan retrieval paths in a memory graph. Key features include Dreaming, an offline consolidation process that merges redundant memories and archives outdated ones, and Feedback, which uses implicit and explicit user corrections to update and refine memory. MindMemOS also supports Skill evolution, where real task trajectories are used to iteratively improve Skills, as demonstrated on SpreadsheetBench-Verified, increasing task success from 51.3% to 57.2% with supervision. On memory benchmarks, MindMemOS achieved 94.03 on LoCoMo and 70.63% on PersonaMem, both state-of-the-art. The system is released under MIT License and provides APIs, SDK, CLI, and plugins for easy integration.