@berryxia: Guys, the MemOS 2.0 open-source project has been updated again! It has gained 9.3K Stars on GitHub ~ This time, 'AI memory' has been upgraded from an advanced clipboard to true 'execute and learn'. Previously, many memory solutions simply stored chat logs and added semantic search, making it look like memory, but it was actually just RAG...
Summary
MemOS 2.0 open-source project update introduces the 'execute and learn' mechanism, enabling the AI Agent to automatically deconstruct and distill experience when completing tasks, evolving hierarchically from raw trajectories to muscle memory, resulting in a dedicated assistant that understands you better as you use it.
View Cached Full Text
Cached at: 05/21/26, 05:38 PM
Folks, the MemOS 2.0 open-source project has been updated again!
It has already earned 9.3K stars on GitHub ~
This time, it upgrades “AI memory” from an advanced clipboard to a true “execute-to-learn” system.
Previously, most memory solutions simply stored chat history and added semantic search — it looked like memory, but was still just RAG in disguise.
The standout feature of the MemOS Local Plugin 2.0 is called “Execute and Learn.”
It doesn’t just remember what you said — when an Agent helps you complete a task, it breaks down the entire execution process into learnable units.
Which step found a key clue, which step was just inefficient probing, which reflection led to later success.
These experiences are automatically distilled into layers:
The bottom layer is raw traces; above that, general patterns; higher still, long-term world models; and at the top, muscle-memory skills.
A dual-feedback mechanism scores automatically — useful patterns get reinforced, inefficient ones fade over time.
Previously, when writing tools with OpenClaw, the code style, naming conventions, and error-handling approaches you refined in the first round would be lost after switching conversations or a couple of days — you’d have to explain everything from scratch.
Now with 2.0, when you start a new task the next day, it directly uses the style you worked out together in the previous session.
This isn’t just “remembering the context” anymore — the Agent, while helping you work, automatically reviews, distills, and evolves.
The more you use it, the better it understands you; the more it feels like your personal assistant.
This update also supports seamless migration for Hermes and OpenClaw — installable with a single command, and the Memory Viewer shows the entire chain of memory clearly.
Pretty interesting — both Hermes and OpenClaw can get it up and running.
Link in comments ~~~
Ears (@RookieRicardoR):
MemOS has made new progress.There are quite a few approaches to AI Memory now, but many still stay at the level of storing chat logs — it looks like memory, but it’s essentially just adding semantic retrieval to markdown.
@MemOS_dev has been working on memory systems for a while, from 1.0 to 2.0. Looking back now, a consensus is gradually becoming clear: the next step that truly determines Agent
Similar Articles
@RookieRicardoR: MemOS has made new progress. There are quite a few AI Memory solutions now, but many are still at the level of storing chat history. It looks like memory, but it's essentially adding a semantic search to markdown. @MemOS_dev has been working on a memory system for a while, from 1.0 all the way...
MemOS Local Plugin 2.0 update has been released, introducing the "Learn by Doing" feature, which allows the agent to convert key steps during task execution into reusable, scorable cognitive assets, thereby achieving continuous learning and memory in the local environment.
@berryxia: Agent memory is incredibly competitive! I have to say, the more people join this track, the better it gets! The Tencent AI team spent a full 6 months tackling just one problem: AI agents frequently dropping context in long conversations. They ended up building a complete memory system and open-sourced it directly. After reading their sharing, my biggest takeaway is...
Tencent AI has open-sourced an Agent memory system that significantly improves token efficiency and agent consistency in long dialogues through three methods: real-time context compression, Mermaid task maps, and Persona memory. Token consumption is reduced by 61%, and persona consistency jumps from 48% to 76%.
@WY_mask: Build persistent memory engine for all kinds of AI coding assistants http://github.com/rohitg00/agentmemory… Silently records code changes and context in the background, automatically extracts and compresses into structured memory, saves Token consumption from long context, associates past information, as…
agentmemory is an open-source tool that provides persistent memory for AI coding assistants. It silently records code changes and context, automatically extracts and compresses them into structured memory, reduces Token consumption, and supports multiple mainstream platforms such as Claude Code and Codex.
@MaxForAI: Yesterday, ByteDance Seed open-sourced a very interesting checkpoint, TaskMem. It is trained on Qwen3-VL-30B-A3B, with the goal not being to directly answer questions, but to enable multimodal Agents to learn to generate more useful long-term memory from video/environment streams. The key is to let the Agent learn in continuous video…
ByteDance Seed has open-sourced the TaskMem checkpoint, trained on Qwen3-VL-30B-A3B. It uses two-stage reinforcement learning to enable multimodal Agents to learn to generate long-term memory from video streams, achieving significant improvements on benchmarks such as VideoMME and EgoLife.
@teach_fireworks: Don't let an AI Agent make the same mistake twice. Recently I came across an interesting open-source project: RoBrain https://github.com/adelinamart/robrain… RoBrain is the "decision memory layer" for AI coding teams, recording each and every...
RoBrain is an open-source shared memory layer for AI coding teams that captures technical decisions, rationale, and rejected alternatives across sessions and tools like Claude Code, Cursor, and Copilot, preventing agents from repeating past mistakes.