What was released
On October 7, 2026, Shanghai startup MemTensor announced open availability of Metis, which it calls the world's first "memory foundation model." Weights are on Hugging Face in three sizes: 4B, 9B and 27B, built on Qwen3.5. The technical paper (arXiv 2607.26760) came out in August with Renmin University, NUS, Shanghai Jiao Tong and Tongji.
"World's first" is the company's own claim; fast weights and in-network memory have decades of research behind them.
The problem
Today AI agent "memory" is almost always external: vector stores, RAG, saved summaries re-inserted into the prompt. It works, but it's an extra stack, costs tokens, and the model never learns what to keep.
How Metis works
- Each full-attention layer gets a Metis block. The Local Memory Block holds a memory matrix that persists across interactions; the Hyper Memory Block has learned parameters that decide what to store and how to read it.
- After each step, the model compresses informative states into memory (Gated Delta Network update).
- At query time, memory attention reads that state and blends it with normal attention, as if history were in the prompt, without paying for the tokens.
- At inference weights stay frozen; only memory changes, via a plain forward pass, no gradients.
Stated limitations
The project calls Metis a prototype:
- Capacity is limited: performance drops when an update exceeds a few hundred words.
- Long trajectories cause interference and semantic confusion.
- Gains were measured mostly on synthetic memory data and no-context QA; real long-running agent use is unproven.
Company context
Per The Yangtzeer, MemTensor raised about 100 million yuan in July (backers include Huawei's Habo Capital and Honor's investment arm) and another round in September, and says it serves 500+ enterprise customers such as Haier and Lenovo. These are company-reported figures.
What it means for businesses
- Native memory is still research. For production, RAG and external memory remain the mature, auditable option.
- Privacy moves. If memory lives in an internal model state, deleting or auditing a specific fact is harder than in a database, which matters under GDPR, LGPD and Japan's APPI.
- Worth a lab test: the 4B and 9B checkpoints run on ordinary GPUs and can be compared against your RAG stack.
- Check the license in the repo before commercial use.
Sources
- GitHub, MemTensor/Metis: https://github.com/MemTensor/Metis
- arXiv 2607.26760 (Metis: Memory Foundation Model): https://huggingface.co/papers/2607.26760
- The Yangtzeer (2026-10-09): https://yangtzeer.com/news/startups/memtensor-memory-native-ai/
- State of AI (Substack): https://stateai.substack.com/p/metis-llm-memory-without-the-cumbersome
Want to evaluate memory options for AI agents (RAG, external memory or new models)? KGA IT offers a free consultation: kga-it.com.