In 1999, a group of students from the University of Science and Technology of China started a business, aiming to build a one-stop self-learning platform for Agents. HSG Seed and Mingshi Capital invested in it.
Text | Zhong Chudi
Editor | Zhou Xinyu
One - sentence Introduction
ACONTEXT is a context data management platform for Agents. By providing data storage, decision - path observation, and agent self - learning services, ACONTEXT provides users with the path and decision - making reasons when Agents execute complex tasks. At the same time, it increases the task execution success rate of Agents by 30% - 50%.
Financing Progress
Recently, ACONTEXT completed a seed - round financing of millions of dollars, jointly invested by Sequoia China Seed Fund and Mingshi Capital.
Team Introduction
Founder and CEO Ye Jianbai (Gus): He once participated in the research and development of the Bing search advertising algorithm at Microsoft Research Asia. Since 2023, he joined the large - model company "Beyond Light - Years" founded by Wang Huiwen and participated in the distributed training of large models. After that, he joined the AI application company "Muyan Zhiyu" founded by Zhang Yueguang, the former person - in - charge of "Miaoyacamera" at Alibaba, mainly responsible for internal model training and the construction of the RAG system.
During his career, Ye Jianbai was responsible for the research and development of several popular open - source projects. For example, the AI long - term memory solution Memobase led by him has obtained 2.6K Stars on GitHub and served several well - known AI products; he once reproduced the GraphRAG algorithm of Microsoft's 20,000 - line code with 1,000 lines of code and obtained 3.7K stars on Github.
The main members of the ACONTEXT team all come from large companies such as ByteDance's Volcengine, Baidu, Microsoft, or star AI companies.
Products and Business
AI applications are undergoing a rapid evolution from Chatbots to Agents. The strategic technology trend report released by Gartner in 2025 shows that Agents have ranked first among the top ten annual technology trends. It is estimated that by 2028, 33% of enterprise - level software will be driven by Agents.
Therefore, how to support the large - scale implementation of Agents has become the most urgent technological proposition in the entire industry. Different from the simple dialogue interaction of Chatbots, the core characteristic of Agents lies in their ability to plan, execute, and reflect.
During his work at the previous company, Ye Jianbai explored different forms of AI products such as ChatBots and Agents. Two data - related trends caught his attention:
On the one hand, the data generated by AI Agents during operation is at least a thousand times that of Chatbots. "One million Tokens can allow ChatBot users to use for about 3 - 7 days, but only allow Coding Agent users to use for 10 - 20 minutes."
On the other hand, currently, there is no solution for the storage, management, and utilization of Agent behavior data on the market. Ye Jianbai told us that regardless of the form of AI applications, the core data composition of their storage is to package user behavior into a data format understandable by large models, that is, Context. For Agents, the in - depth mining and utilization of Context determine the upper limit of Agent intelligence and user experience.
These two trends made Ye Jianbai decide to use Context as the starting point for entrepreneurship: to build a set of infrastructure for the storage, monitoring, and self - learning of Context data to explore the utilization value in Context.
At the beginning of 2025, Ye Jianbai made his first product attempt - Memobase. This is a memory solution targeting consumer - level Chatbots. AI applications connected to this solution can form user profiles based on users' historical data within 100 milliseconds. Based on the user profiles obtained from Memobase, Chatbots can generate personalized responses and improve user experience.
This set of solutions has been implemented in more than 10 AI startups such as Zao Meng Ci Yuan and Nie Ta, providing memory modeling services for more than 50,000 active users every day.
However, Ye Jianbai judged that the barriers of a single - point memory storage solution are limited. "If you don't control the Context data and the data is stored in third - party cloud providers, a single - point memory solution is easily 'annexed' by upstream providers." At the same time, at the commercialization level, it is difficult to measure the ROI (return on investment) of Memobase, which is not conducive to the company's formulation of a charging model.
In his opinion, the core barrier of middle - layer providers lies in building a developer ecosystem. The logic behind this is that for middle - layer providers, Context is the most valuable data asset. Therefore, to accumulate more Context assets, a prosperous developer ecosystem needs to be established.
In November 2025, the team launched a new product, ACONTEXT. Essentially, ACONTEXT is a Context data management platform for Agents. By precipitating, understanding, and mining Context data, it reduces the threshold for the development, implementation, and management of Agents and improves the efficiency of Agent development and management.
The services provided by ACONTEXT for Agent developers are mainly implemented in three stages:
First, in the Agent development stage, ACONTEXT solves the complex "underlying infrastructure" problems in the early stage. It builds pipelines for data storage and use around context data, making Agent data storage ready - to - use.
This approach preserves the integrity of the global reasoning of large models. Developers no longer need to consider issues such as the storage of multi - modal information, the allocation of sandbox environments, and the compatibility of file systems, nor do they need to write thousands of lines of code to connect to various scattered databases, shortening the product launch cycle.
Second, in the Agent launch stage, ACONTEXT provides real - time monitoring and management services for Agents.
Traditional feedback relies on inefficient user surveys or manual retrieval in databases. ACONTEXT has an auditing Agent built into the background. It will automatically disassemble Agent and user behaviors, providing developers with the specific execution process of Agents and real - time user feedback.
Finally, ACONTEXT improves the stability of Agents by "establishing Agent Skills". The auditing Agent is used for both monitoring Agent behaviors and establishing Agent Skills. ACONTEXT's self - learning system will extract the successfully executed paths and turn them into the Agent's exclusive memory or skill packages. It will also analyze failed tasks and extract experience.
This means that when users put forward similar requirements next time, Agents no longer need to explore blindly but can directly call the verified processes, reducing the uncertainty of Agent behaviors.
Ye Jianbai told us, "Our goal is to help customers achieve the Context data flywheel. Only in this way can Agent products start to build user stickiness and product barriers."
During the POC (proof of concept) stage of ACONTEXT, this solution helped Agents increase the task success rate by 30% - 50% and reduce the number of operation steps by 10% - 30%.
Founder's Thoughts
1. Different from the complex data mentality in the past Internet era, the data form of AI applications is extremely single.
In the Web 2.0 era (such as Taobao and Meituan), developers need to process hundreds of scattered data information such as user clicks, stay time, and shopping carts. But in the AI era, 99% of the data is ultimately in one format, that is, Context. Products that cannot convert their own data into a format convenient for putting into the context of large models will face great challenges.
2. Currently, the entire industry is advocating Embedding (vector retrieval) to solve the memory problem, but this is actually the key to limiting the intelligence of large models.
Large models have strong text - processing capabilities, with the number of parameters reaching the scale of hundreds of billions. However, vector models are very small, usually with only dozens of MB of parameters. If vector retrieval is used, the retrieved content is likely to be out of context. Therefore, the real evolution direction is to return the decision - making power to large models and let them decide which part of the information to dig deeper. Only in this way can Agents show combinatorial reasoning abilities.
3. Integrated innovation is the main battlefield for young people.
Single - point infrastructure such as computing power cluster management and distributed training emphasizes scale and assets, and large companies have more advantages in this regard. Compared with single - point infrastructure, integrated infrastructure does not mean inventing a new database form but rather combining and optimizing third - party resources and services to output new product and service forms, which tests the early - stage team's choice of ecological niche.
Large companies will have a natural path dependence when doing integrated infrastructure. They prefer to sell cloud resources and S3 storage. Young teams have no burdens, and we are more aware of what Agent developers need. Therefore, the opportunity for integrated innovation infrastructure is reserved for young teams.
4. The most noteworthy trend in 2026 is the Environment of Agents.
In 2026, the algorithms and logic of Agents will become highly homogeneous. At that time, the deciding factor will no longer be how smart an Agent is, but the depth of its connection with the Environment.
The Coding Agent can explode first because the working environment of programmers is standard and tidy. But future financial, legal, and government Affairs Agents will face an extremely chaotic data environment without unified standards. Therefore, in 2026, the Environment of Agents should be a frequently mentioned field.
5. For Agent infrastructure, the core barrier lies in the ecosystem.
For Agent infrastructure, the real barrier does not lie in technology or data but in the ecosystem, that is, your users are willing to create content for your product. For example, when people search for "data context data" or "context data platform", if 6 out of the top 10 results on Google are related to you, then this barrier is truly formed.
Welcome to communicate!