Memory is evolving into an independent track.
In the Agent era, large models are no longer competing solely on single-turn Q&A capabilities.
Writing code, parsing materials, calling tools—the list of tasks AI can handle keeps growing. However, when Agents take over continuous projects spanning multiple turns, challenges such as inability to maintain historical communication, difficulty in reusing past experiences, and skyrocketing computing power costs for long conversations arise one after another. Memory has become an unavoidable hurdle standing in the way of long-term Agent deployment.
On October 20, MemTensor (Shanghai) Technology Co., Ltd. officially released and open-sourced its memory-native foundation model series, MemT 1.0, including Flash and Nano versions. Unlike general solutions that extend context windows or attach external knowledge bases, this Shanghai team chose to modify the internal architecture of the model, allowing historical information and task experience to directly participate in subsequent computations.
Ten days later, MemT1.0 cumulative downloads surpassed 100,000, averaging over 10,000 per day; on the first day of the technical report's release, it ranked second on Hugging Face's (the world's leading open-source AI model community) daily paper leaderboard, and Hugging Face officials also published articles analyzing its memory mechanism and real-world performance; numerous overseas developers have already begun secondary development.
What exactly is new about this solution, and why is it worth trying for so many developers?
Embedding Memory into the Model: Not Flipping Through Notes, But Remembering It Directly
The industry's mainstream approaches to handling memory generally fall into three categories.
One approach is extending the context window, which is equivalent to bringing the full set of historical minutes to every meeting. The materials are complete, but as the conversation lengthens, the content the model needs to process expands accordingly, and costs rise with the length of the dialogue.
Another approach is training knowledge into the model itself, allowing direct invocation. However, once a single piece of memory needs to be modified, it often requires adjusting the entire model, lacking flexibility.
A third approach is attaching an external knowledge base for retrieval, namely RAG. It is like setting up a filing cabinet where data can be queried and modified, but the model can only retrieve external materials and cannot automatically digest the experience accumulated from interactions.
These methods each have their uses. MemTensor further explores: Can new information from interactions be continuously written into the model internally, read and used directly subsequently, without needing to retrain the model for every interaction?
MemT1.0's approach is to carve out an updatable "note area" inside the model. After each round of conversation, new information is automatically written in; during the next response, it participates in thinking directly alongside the current question, without needing to re-read all original conversation records. The storage output is a memory snapshot specific to the model. Humans cannot directly read the text within the snapshot, but it supports saving, loading, and distribution, allowing the model to invoke it directly.
The team named this paradigm "memory-native"—memory is no longer an external tool but a component of the model's computational flow itself. Previously, one would flip through a notebook before working; now, things are remembered in the brain. Among the two open-sourced versions, Nano has 36 billion parameters, and Flash has 288 billion parameters.
Let's feel it through a real test scenario: An exhibition was originally scheduled for Saturday at 3 PM, the official notice changed it to Sunday at 4 PM, then someone proposed changing it to Monday, but it wasn't finalized. Afterwards, the topic shifted elsewhere, and they chatted for more than 500 rounds. When asked again about when the exhibition was originally scheduled and when it was finally changed to, Nano answered correctly, clearly distinguishing between the original arrangement, the official change, and the unadopted suggestion, without re-reading the chat history during the response.
In real project collaboration, the value of this capability is particularly prominent—proposals, revisions, and invalid information are intertwined within projects, and AI needs to discern the validity of information, distinguishing which are merely alternatives and which are final confirmed conclusions.
In 11 public memory evaluations, the average scores of the two models were 50.6 and 47.0 respectively, ranking in the top two in the parametric memory track, exceeding the previous best baseline by 12.5 points and 8.9 points respectively.
Remembering facts that have occurred is just the basics; greater value lies in reusing work experience. The team distilled 1,040 debugging experiences from over 40,000 real programming trajectories, which can be written into internal memory when AI tasks hit bottlenecks.
MemT 1.0 Memory Performance in Benchmark Tests
Controlled experiments showed: Without loaded experience, AI locates code issues and starts modifying directly; after injecting the experience of "first reproduce the problem, then implement modifications, and conduct verification upon completion," the model first writes tests to reproduce the problem, actively reviews boundary conditions after modification, and then iteratively corrects them.
In SWE-bench Verified, consisting of 500 code repair tests, the proportion of tasks successfully executed three times consecutively increased from 54% to 62.8%, and the average task success rate rose from 66.07% to 69.73%. For enterprises, stable reproduction of results is far more practically meaningful than accidentally answering correctly, as it effectively reduces the burden of manual review. Although the experience in current experiments still needs to be constructed externally, retrieved, and written in, the results demonstrate that existing methods can influence subsequent model execution through internal memory.
In terms of computing overhead, MemT relies on fixed-size memory states to avoid the text carrying burden caused by longer dialogues. Controlled tests show that as history length increases from128K to 256K, inference speed reaches 1.86 to 3.99 times that of traditional full-context solutions, storage overhead is only 1/66 to 1/130 of traditional solutions, and the wait time from question to start of answer remains stable around 0.1 seconds. However, the advantage is not obvious in short dialogues; long conversations are its main arena.
Efficiency evaluation results: End-to-end latency and query first-token latency
Dual-Line Product Layout: Capital Betting on the Long-Term Memory Track
Behind MemT is a continuously advancing research roadmap.
In 2023, founder and CEO Xiong Feiyu led the establishment of the Large Model Center at the Shanghai Algorithm Innovation Research Institute, making long-term memory the core direction for the next generation of model architectures. In November 2024, the team founded MemTensor, driving the evolution of scientific research results into product implementation.
Xiong Feiyu graduated from Huazhong University of Science and Technology for his undergraduate degree and obtained a PhD from Drexel University in the United States. With an Alibaba background, he served as the head of Data Intelligence for Alibaba's Business Middle Platform and the Data Platform for Taobao Tmall Group. He has done research and experienced the practical combat of large-scale business systems.
Before the emergence of the MemT model, the team had already launched the memory operating system MemOS.
Enterprise information is often scattered everywhere: Project documents contain background, chat records contain requirements, and execution logs contain processing procedures. What should be kept, how to update changes, and how to quickly locate items when invoking require unified management.
MemT 1.0 Memory-Native Architecture
MemOS and MemT 1.0 belong to two parallel R&D routes with the same goal but clear division of labor, supporting combined deployment.
· MemOS, System-Layer Product: Focuses on managing memory materials. Responsible for storage, retrieval, version control, and scheduling. This product has already achieved industrial implementation, with GitHub Stars exceeding 10,000 for the open-source project, cloud service monthly calls surpassing 50 million, and serving over 500 enterprise customers. It continuously gains frontline pain points and engineering experience from a large number of real business scenarios.
· MemT 1.0, Model Architecture-Layer Product: Internally modifies inference logic, allowing memory to directly participate in model computation, exploring how to make the model digest interaction history and reuse it directly in subsequent tasks.
In other words, MemOS manages external material archives; MemT attempts to let the model grow its own memory. Content requiring precise verification, such as contracts, amounts, and legal provisions, still needs to retrieve the original text; behavioral logic and communication consensus accumulated from long-term collaboration are more suitable for native memory.
Capital has already sensed the potential of this route. In July this year, MemTensor completed a Pre-A funding round of CNY 100 million, with Huawei Hubble, Honor Strategic Investment, SenseTime Guoxiang Capital, Shenzhen Venture Capital, and Heyu Capital joining; in September, it secured another Pre-A+ funding round, led by Oriental Fortune Capital, with additional investment from Shanghai State-owned Assets Investment Futeng Capital.
Two funding rounds in just two months saw industrial capital, state-owned platforms, and market-oriented VCs betting heavily. Funds will be invested in memory-native foundation model R&D, MemOS ecosystem construction, and global talent recruitment. The open-sourcing of MemT 1.0 is precisely the phased achievement of this route reaching the model layer.
Memory Track Heating Up: Placing Pieces Deep Within the Model
Memory is growing from a supplementary function of models into an independent track.
iResearch predicts that China's agent memory market size will grow from CNY 1.44 billion in 2025 to CNY 64.25 billion in 2030, accounting for nearly 26% of the agent infrastructure share, becoming the second-largest independent market after models.
Behind the market growth expectations is the major trend of AI deeply embedding into enterprise business processes. A large amount of Know-how, troubleshooting experience, customer service methods, and pitfalls encountered by teams are precipitated in years of practice and should be continued and reused.
As the track heats up, players naturally increase. Cloud vendors leverage storage resources to deploy memory management, independent platforms focus on general memory middleware, and research such as Google Titans explores internal model memory. However, the vast majority of solutions on the market are still built outside the model.
The route of internally modifying model architecture is rarely traveled and faces many realistic challenges. MemT 1.0 is currently a preview version, not the final solution: Due to fixed memory capacity limits, some details may suffer loss after high-frequency updates; in scenarios requiring precise restoration of massive original text details, external retrieval solutions still hold advantages; how to filter and select between conflicting old and new information, and how to achieve autonomous model experience extraction, remain topics to be conquered.
Precisely because of this, the team chose to open-source the model and technical report, handing it over to global developers for joint verification and iteration.
In addition to Hugging Face and GitHub, the model was simultaneously launched on the ModelScope community; according to company disclosures, among large-scale innovative architecture open-source models, its download data leads. US developers have already packaged Nano into a "persistent memory model" runnable locally, capable of compressing entire conversation histories into approximately 40MB memory archives, aligning with the design of fixed memory states. The deployment threshold is not high either: Official inference guides indicate that Nano requires two graphics cards, and Flash requires eight.
Domestic developers have also completed real-world validation: Running Nano on a single high-end graphics card, injecting four internal coding standards for payment systems into the model to generate a memory snapshot; then starting a new blank session without importing any chat records, loading only the snapshot, and asking it to write a payment callback interface. The model invoked the team's unique component names and error codes; while the control group without memory only output generic code.
However, remembering key points does not equal strict execution; details such as locking syntax and lock names still have deviations, and outputs still require manual review. But this snapshot can be saved, restored, and distributed, providing a new carrier for long-term accumulations such as team standards and troubleshooting experience.
The essence of competition for open-source models is ecosystem competition: The more cases of being downloaded, reproduced, and secondary developed, the more likely the technical route is to precipitate into an industry-standard foundation.
In the Agent era, enterprises need intelligent assistants capable of continuing long-term tasks. Whether memory can be retained and experience reused directly impacts the upper limit of human-machine collaboration. By letting experience participate in computation and memory become capability, this Shanghai tech innovation team has played a "memory-native" foundational card for the next stage of the large model industry.
Text | Nan Yi Edited | Yi Da