AI不再用完即忘：华为诺亚开源MindMemOS，记忆和Skill一起进化
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Huawei's Noah's Ark Lab open-sourced MindMemOS (MIT license), a portable, self-evolving memory operating layer that decouples memory from individual agents and co-evolves stored memories with executable Skills. It matters because it reframes agent memory as reusable long-term infrastructure rather than per-session cache.
- Memory is modeled on an entity–property–time structure that keeps both current facts and full evolution trajectories, with two generation paths (MindVanilla for open-domain, MindSchema for domain-specified entities/properties) and a "Compact Search" agentic retrieval layer doing multi-hop and timeline traversal.
- Offline "Dreaming" consolidation merges duplicates and archives superseded facts via supersedes links: on MemoryAgentBench FactConsolidation it archived 19.4–23.5% of active memories while raising QA accuracy by up to 10.3 points.
- Feedback mining turns explicit and implicit user corrections into add/update/archive/delete/reinforce operations, updating semantic structure rather than just patching one record (PersonaMem-Evo case study).
- Skill evolution from real execution traces lifted SpreadsheetBench-Verified success from 51.3% (unsupervised) to 57.2%±2.4% (supervised), +9.2 points over the un-evolved initial skill — which itself scored below no-skill. Reported memory benchmarks: LoCoMo 94.03, PersonaMem 70.63%. Access via FastAPI, Python SDK, CLI, and an OpenClaw plugin; full technical report still pending.
Sources (1)
AI不再用完即忘：华为诺亚开源MindMemOS，记忆和Skill一起进化
TL;DR - Huawei's Noah's Ark Lab open-sourced MindMemOS (MIT license), a portable, self-evolving memory operating layer that decouples memory from individual agents and co-evolves stored memories with executable Skills. It matters because it reframes agent memory as reusable long-term infrastructure rather than per-session cache.
- Memory is modeled on an entity–property–time structure that keeps both current facts and full evolution trajectories, with two generation paths (MindVanilla for open-domain, MindSchema for domain-specified entities/properties) and a "Compact Search" agentic retrieval layer doing multi-hop and timeline traversal.
- Offline "Dreaming" consolidation merges duplicates and archives superseded facts via supersedes links: on MemoryAgentBench FactConsolidation it archived 19.4–23.5% of active memories while raising QA accuracy by up to 10.3 points.
- Feedback mining turns explicit and implicit user corrections into add/update/archive/delete/reinforce operations, updating semantic structure rather than just patching one record (PersonaMem-Evo case study).
- Skill evolution from real execution traces lifted SpreadsheetBench-Verified success from 51.3% (unsupervised) to 57.2%±2.4% (supervised), +9.2 points over the un-evolved initial skill — which itself scored below no-skill. Reported memory benchmarks: LoCoMo 94.03, PersonaMem 70.63%. Access via FastAPI, Python SDK, CLI, and an OpenClaw plugin; full technical report still pending.