Switch agents or even just start a new session, and it's like getting a brand-new employee who knows nothing about you — user preferences, project context, lessons learned from past mistakes, all wiped clean. You start over from scratch, "training" the new agent all over again. It's the most frustrating pain point in today's Agent experience.

Huawei Noah's Ark Lab's open-source MindMemOS aims to solve this at the root. It's not another RAG scheme — it's a transferable, self-evolving memory operating layer for AI Agents that upgrades memory from an Agent's private cache to an independent asset reusable across applications and frameworks.

3D Memory Structure: Entity, Property, Time

Traditional Agent memory stores memories as text chunks or vector blocks, retrieving via similarity search. MindMemOS completely reorganizes memory around a 3D coordinate system for open-world information: Entity for persistent objects like people, projects, files, tools; Property for facts, preferences, and states of objects; Time for when properties hold and when they change.

"User preferred X before" and "User switched to Y now" are no longer disconnected text chunks — they're two states of the same entity-property pair on a timeline. At retrieval time, the system doesn't look for "the paragraph most similar to the query" — it searches the memory graph for the shortest path that fully covers the needed information.

Most intriguing is the Dreaming mechanism. Humans consolidate memories during sleep — merging duplicates, retiring outdated facts, forming abstractions from experience. MindMemOS's Dreaming module does the same during Agent idle time: resolves conflicting memories, archives stale information, moves roughly one-fifth of active memories to archive (not proactively recalled), and paradoxically improves accuracy across multiple tests. On MemoryAgentBench, Dreaming compressed 19.4-23.5% of active memories while improving QA accuracy by up to 10.3 percentage points — less is more; reducing interference is itself a quality upgrade.

Skills Evolve Too: From Static Prompts to Dynamic Capabilities

If factual memory makes Agents understand users better, Skill evolution determines whether they can turn experience into capability. MindMemOS uncovered a counterintuitive finding: unoptimized initial Skills can be worse than no Skill at all. On SpreadsheetBench, giving agents an openpyxl/pandas usage guide as an initial Skill actually produced lower success rates than the No-skill baseline — outdated, redundant rules created interference.

MindMemOS evolves Skill rules by collecting real task trajectories and learning from both successful paths and failure patterns in an unsupervised manner. For instance, it learns from failures: "Don't treat formula text as values," "Build a stable snapshot of the source sheet before filtering/sorting," "After saving, you must reopen and reverify." These aren't extra API tutorials — they're compressed lessons from real error histories. Task success rate ultimately improved from 51.3% to 57.2%.

On benchmarks, MindMemOS achieved 94.03 SOTA on LoCoMo long-conversation memory and 70.63% on PersonaMem personalized memory. The project is MIT-licensed with FastAPI endpoints, Python SDK, CLI, and OpenClaw plugin. A cloud service at mindmemos.cn is open for trials.

The gap in long-term Agent intelligence was never about how well one task goes — it's about what remains after a hundred tasks. MindMemOS's answer is a memory system that tracks changes, proactively organizes, accepts corrections, and continuously evolves. It may be the essential infrastructure for Agents to graduate from "disposable tools" to "long-term collaborators."

See you tomorrow.