- MemTensor’s Metis model embeds persistent memory into the model itself, aiming to help AI systems learn from experience over time
- The startup has raised two funding rounds in two months and says it serves more than 500 enterprise customers
Shanghai-based AI memory startup MemTensor (记忆张量) has open-sourced what it describes as the world’s first memory-native foundation model, a system designed to let AI retain and build on past experience rather than treat each interaction as a fresh start.
The company announced on October 7 that its Metis model was available on Hugging Face in three sizes — 4 billion, 9 billion and 27 billion parameters. Downloads exceeded 10,000 in the first 30 days, according to MemTensor.
Built-in memory
Most large language models struggle to retain information across interactions. Longer context windows allow them to process more conversation history, while retrieval-augmented generation (RAG) lets them retrieve information from external databases.
But both approaches rely on supplying relevant information when needed, rather than giving the model a persistent internal memory.
Metis takes a different approach by maintaining a persistent memory state within the model. It compresses historical information into an internal memory matrix that the model can retrieve during inference and update with new experience, without requiring retraining for every change.
The goal is to move AI beyond what MemTensor calls “one-shot intelligence” toward systems that can develop a more durable understanding of users and tasks through repeated use.
Three stages of research
The release builds on three stages of research. In 2023, the team proposed its Memory³ hierarchical memory theory to address how AI memory should be modeled.
From 2024 to 2025, it developed MemOS, which the company describes as the first operating system for large-model memory, designed to manage how information is stored and retrieved.
MemTensor says MemOS has performed strongly in evaluations of agent memory and user memory. Its open-source project has surpassed 10,000 GitHub stars, while its cloud service handles more than 50 million calls a month, according to the company.
Metis moves the work another step forward by integrating memory into the model itself.
MemTensor has completed two funding rounds in the past two months. In July, it raised around 100 million yuan in a pre-Series A round backed by Huawei’s Habo Capital, Honor’s investment arm, SenseCap, Shenzhen Capital Group and MSA Capital. In September, it secured a pre-Series A+ round led by Oriental Fortune Capital.
Investors bet on AI memory
The company says it serves more than 500 enterprise customers, including Haier and Lenovo.
Founder Xiong Feiyu (熊飞宇) previously led data intelligence for Alibaba’s business middle platform, where he oversaw the development of a digital commerce knowledge graph at a scale of hundreds of billions of data points.
He founded MemTensor in November 2024 with a team from the large-model center of the Institute for Advanced Algorithms Research, Shanghai. The core team has an average age of under 30, and chief scientist Yang Hongkang (杨泓康) is a Princeton University graduate.
Why this matters
Persistent memory could help AI agents carry context across longer tasks and reduce the need to repeatedly provide background information.
But a model’s ability to retain useful information over time also raises practical questions about memory accuracy, privacy and how reliably it distinguishes lasting preferences from outdated or irrelevant details.
Metis’ performance in real-world, long-running applications will be an important measure of whether memory-native design can deliver on its promise.
Header image credit: MemTensor official website
