AIBRIEF
MindMemOS, from a research lab inside Huawei, runs an offline consolidation pass before the AI assistant recalls anything; the published benchmark numbers are self reported.
Every new chat with an AI assistant starts from zero. Preferences, the corrections made last week, the project you were working on: none of it carries over, even when you stay on the same service. MindMemOS, an open-source project released this week by Huawei Noah's Ark Lab, a research lab inside Huawei, is one specific attempt to fix that.
The fix is not a bigger context window. MindMemOS gives AI agents (the software assistants that take multi-step actions on a user's behalf) their own memory layer. The system stores what the project calls an "Entity-Property-Time" structure: the current state of a fact, how it changed, and its relations to other facts. A separate offline "Dreaming" pass then merges duplicates, archives contradictions, and promotes reliable memories before the agent recalls anything. A second module, Skill Evolution, distills repeated behavior into reusable skills. Huawei reports a 57.2% task success rate on SpreadsheetBench-Verified.
The benchmark numbers, on the project site and in QbitAI coverage, are self-reported: 94.03 on LoCoMo (long-conversation memory) and 70.63 on PersonaMem. On MemoryAgentBench's FactConsolidation task, the Dreaming pass is said to compress roughly a fifth of active memory while lifting QA accuracy by up to 10.3 percentage points. The GitHub repository is public.
What remains unknown: independent reproduction of the Dreaming numbers, comparisons to existing agent-memory projects like Mem0 or Letta, and any signal of whether a portable memory layer actually changes how an AI assistant feels in everyday use.