Metis🚀 Memory Foundation Model released by Memtensor Research Group
Probably the first LLM that internalizes memory into its backbone👀
- 4B/9B/27B weights + code + paper
- No external RAG: memory is inside the model
- One forward pass = full read/write
- Weights frozen at inference: deploys like a standard LLM
New paper's here: Metis — the 1st prototype of a memory foundation model.
In Metis, the model weights remain fixed during inference. What changes instead is a persistent memory state.
This state carries information across interactions, updates during the forward pass, and is retrieved through memory attention when relevant — all without gradient updates.
We refer to this architecture as a memory foundation model. Metis is created to explore what becomes possible when memory is built into the model's computation, rather than added as an external system around it.
Paper, code, and checkpoints:
https://t.co/jOllnXFhSw
If this direction resonates with you, we'd appreciate a 🗳️ on Hugging Face.
🏆 Siiiiuuuuuu!!! MemOS FINALLY reached 10K Star Goal!
11 months ago, we open-sourced MemOS with one goal: making sure AI never loses its memory, and can keep evolving on its own 🧑🏻🚀
We started gaining real traction in the dev community, and once we crossed 8K stars, something shifted — contributions exploded. From March to June alone, we received 386 issues and 125 PRs. That's not just more people using the tool; it's more people willing to roll up their sleeves and help build it with us 😎
A heartfelt thank-you to everyone who's contributed — whether you joined last week or you were here when we had just 1 star 💕
10,000 stars is a huge milestone, but what excites us more is how the community has evolved. Some of you are fixing core bugs. Others are doing secondary development or building out deployment solutions. And external products — like Tavily — are now reaching out to integrate with us directly.
MemOS isn't just being noticed anymore. It's being used, and used well 📈
Appreciate everybody through issues, docs, or PRs, MemOS is getting better because of all of you 🧠✨
Thanks for Adina 🙏
One thing worth adding: MemPrivacy is built specifically
for the conversational reversibility loop — sensitive content gets desensitized on-device before leaving, and the original values are restored locally when the cloud reply comes back.
This makes it a fit for any edge-side scenario where private data shouldn't leave the device but the cloud agent still needs semantic context to respond well — health assistants, finance copilots, on-device personal agents.
MemPrivacy ㊙️ a lightweight privacy preserving model for edge cloud AI agents from MemTensor.
Instead of destroying context with masking, it preserves semantic structure using typed placeholders like <Phone_Number_1> 👀
✨ 1.7B/4B - RL/SFT
✨ High precision privacy extraction
✨ 4 level privacy taxonomy (PL1–PL4)
✨ Semantic preserving typed placeholders