MemTensor, the Shanghai AI developer founded by former Alibaba data executive Xiong Feiyu, released a family of models designed to maintain memory inside the model backbone instead of repeatedly retrieving old text from an external database.
The Metis release, submitted as a research paper on July 29th, includes models labeled 4B, 9B and 27B. The Hugging Face collection lists their full sizes at roughly 5 billion, 9 billion and 28 billion parameters, respectively, and says the family is based on Qwen3.5. Adina Yakup (@AdinaYakup) highlighted the release in a July 31st thread on X.
Metis is the model-level expression of the thesis Xiong has pursued since forming MemTensor in November 2024: long-running AI agents need state that persists across interactions. Xiong previously led data intelligence for Alibaba's business platform and the data platform serving Taobao and Tmall, according to 36Kr's reporting on MemTensor. He established the Large Model Center at the Shanghai Algorithm Innovation Research Institute in July 2023 before spinning the group into MemTensor.
"Memory is becoming the new infrastructure in the Agent era," Xiong told 36Kr.
Memory without replaying the transcript
Most agent memory systems save information outside the language model, retrieve selected records and insert them into a later prompt. That approach gives developers an inspectable store, though each query can require retrieval, ranking and another pass through the selected text.
Metis instead inserts memory blocks into Transformer layers. A local memory block maintains a dynamic matrix across interaction steps. A second component selects information from hidden states, updates that matrix and later reads from it through a dedicated memory-attention path. The model's learned weights remain frozen during inference; the memory state changes through forward computation without gradient updates.
In practical terms, a developer can give Metis information in one interaction, commit that exchange to the model's memory state and ask about it later without resending the original text. The state has a fixed size, so the cost of querying it does not grow directly with the full transcript.
The authors describe Metis as the first prototype of a "memory foundation model," a category they define as models with persistent internal state and learned procedures for remembering, updating and forgetting. That first claim belongs to the researchers. Earlier work has explored recurrent state, test-time adaptation, parametric memory and memory-augmented neural networks, and the paper spends considerable space distinguishing Metis from those approaches.
A research preview with visible limits
MemTensor's own release draws a firm boundary around what has shipped. The GitHub repository includes the architecture, inference examples, a training and evaluation harness, code and official checkpoints. It omits the training data. The repository code uses the PolyForm Noncommercial License, which prohibits commercial use without a separate agreement. The repository also says model weights and datasets are governed separately from that software license.
The paper reports that Metis performs strongly on the researchers' memory-operation evaluations and transfers to out-of-distribution memory tasks. Those results have yet to accumulate broad independent replication. The authors also document the central tradeoff: compressing history into a fixed-size state loses information as tasks grow longer. They report occasional confusion when multiple pieces of information blend in the latent memory space, and their experiments show that access to complete source text remains the strongest setting on several memory operations.
That makes Metis an architecture experiment rather than a finished substitute for retrieval-augmented generation. MemTensor's roadmap explicitly keeps hybrid systems in view, combining internal state with external stores when exact records, long histories or auditable source material matter.
MemTensor is pursuing memory at two layers
Metis follows MemOS, MemTensor's separate framework for storing, retrieving and governing external memory for models and agents. The two products give Xiong parallel routes into the same market. MemOS can sit around existing models as infrastructure, while Metis asks developers to adopt a model architecture trained around native memory.
That strategy has attracted hardware and device-linked capital. In July, MemTensor completed a 100 million yuan Pre-A round from Huawei's Hubble Investment, Honor Strategic Investment, SenseTime Guoxiang Capital, Shenzhen Venture Capital and Yu Capital, according to 36Kr. An earlier angel round of nearly 100 million yuan included Fusion Capital, Suanfeng Information and CICC Capital. MemTensor has not published a post-money valuation.
The investor mix fits the technical bet. Devices and persistent agents cannot afford to replay an expanding conversation every time they act. A compact internal state could lower repeated context processing and preserve continuity on hardware with tighter compute and latency limits. Metis still has to prove that its compressed state can remain accurate, controllable and recoverable over the long workflows where memory matters most.