Developing memory capabilities for large language models, Hao Jianye has closed three rounds of financing within half a year.

If a person is praised as "smart" or "quick-witted", he often possesses two core capabilities: one is a good memory that enables him to retain past information, and the other is strong comprehension that allows him to form judgments based on the information he remembers.

This logic also applies to AI. Unlike traditional AI that processes tasks from scratch every time, AI with memory capability can continuously retain users' preferences, historical interactions and behavior habits, form a more stable understanding of users through repeated usage, and adjust subsequent responses and actions accordingly. For AI agents that require long-term collaboration, memory is becoming an infrastructure that determines product experience and user stickiness.

Capital has already placed its bets. Recently, Shenzhen MemoraX Technology Co., Ltd. (hereinafter referred to as "MemoraX AI"), a company focused on R&D of AI memory infrastructure, announced the completion of a seed++ round of financing worth hundreds of millions of yuan. This round of financing is jointly led by Shangshi Capital, Haitang Capital and Ren'ai Capital, with multiple first-tier investment institutions participating in follow-on investment, and existing shareholders continuing to increase their stakes. The company's financing pace is quite intensive: it completed the seed round in April and the seed+ round in May. In less than half a year, it has completed three rounds of financing with a total amount of hundreds of millions of yuan.

The reason for capital's intensive investment first lies in the recognition of the team itself. CEO Hao Jianye is a rare entrepreneur in the industry who has both top academic background and large-scale industrial implementation experience. Back in April, when the news of the company's registration was just released, many investors had already taken the initiative to contact the team.

But more important than the team is that capital is re-evaluating the value of "memory". In the second half of 2026, AI memory has gradually evolved from a technical concept to an independent industrial track. According to third-party estimates, the global market size of Agentic AI orchestration and memory systems is expected to grow from about 439 billion yuan in 2027 to about 1.992 trillion yuan in 2030, with a compound annual growth rate of about 35.3%.

A new story is unfolding in the next round of competition for large models.

Investors Rush to Invest in Huawei's Former Chief Expert

The AI industry is full of top talents, but entrepreneurs with top-tier experience in industry, academia and research like Hao Jianye are very rare.

Dr. Hao Jianye graduated from the Chinese University of Hong Kong, supervised by Prof. Ho-fung Leung, with his research focused on game intelligence. After graduation, he went to the Computer Science and Artificial Intelligence Laboratory of the Massachusetts Institute of Technology and the Singapore University of Technology and Design successively to serve as a postdoctoral researcher.

In 2015, Hao Jianye returned to China and joined Tianjin University. When AlphaGo was launched and sparked a new upsurge in reinforcement learning, relevant Chinese communities and materials in China were still very limited at that time. Hao Jianye led the establishment of the Deep Reinforcement Learning Laboratory at Tianjin University, which became one of the earliest domestic scientific research teams to systematically carry out research in this direction.

After that, he was successively promoted to Elite Professor of Tianjin University, recipient of the National Science Foundation for Excellent Young Scholars of China, and selected into the list of the top 2% scientists in the world. In the past two years, his output of research papers at the top three conferences ICML, NeurIPS and ICLR ranks top 10 globally, with more than 15,000 citations on Google Scholar.

Compared with his pure academic resume, what makes Hao Jianye more distinctive is that he has almost fully experienced several key stages of AI moving from the laboratory to the industry.

Over the past years, he has been deeply engaged in reinforcement learning and completed a large number of industrial implementations around deep learning technology, with application scenarios covering autonomous driving, Huawei's ICT product lines, supply chain and chip design and other fields. Before joining MemoraX AI, he successively served as Director of Huawei Decision Reasoning Laboratory, Director of Huawei Large Model Algorithm Laboratory, Technical President of Huawei Medical Corps, and Chief Expert of Huawei's Decision Intelligence direction.

During his tenure at Huawei, Hao Jianye led the incubation of multiple industrial projects from 0 to 1, creating hundreds of billions of yuan in economic benefits. Entering the era of large models, he shifted his focus to post-training of large models, taking charge of work including reinforcement learning for large models, post-training and industry model applications.

At the end of November last year, Hao Jianye left Huawei. It is understood that as early as the second half of last year, he had begun to seriously consider starting a business. What made him make up his mind was a seemingly simple but very difficult question to answer: when the basic model system of large models has gradually matured, what on earth is the core technology of the next generation of large models?

Past industrial experience gave him the answer — memory.

In Hao Jianye's view, the foundation model itself mainly solves the problem of one-time calculation and reasoning, but when large models are truly applied in all walks of life, the interaction between users and agents will not be one-off. As interactions continue to occur, users' historical behaviors, preferences, task contexts and industry knowledge will continue to accumulate, and the model will face increasingly long contexts. The longer the context, the higher the cost and computational pressure. How to effectively store, retrieve, understand a large amount of historical information and call it back at the right time has also become a new proposition beyond the foundation model.

Hao Jianye believes that memory is not a nice-to-have feature, but is becoming a necessary basic capability for agents to truly achieve long-term interaction and continuous evolution.

From External Plug-in Memory to "Model Endogenous Memory"

Since the beginning of this year, the memory capability of large models has attracted widespread attention, and capital has already made relevant layouts.

Dnotitia, a South Korean deep tech company, is one of them. In April this year, the company focused on AI memory chips completed a $61.2 million Series A financing. Its core product VDPU (Vector Data Processing Unit) adopts In-Memory Computing technology, which directly performs matrix operations inside the memory array, reducing the overhead caused by data transmission at the hardware level.

But in Hao Jianye's view, AI "memory" cannot be simply equated with information storage. Storage is only the underlying carrier of memory, which can be files, knowledge graphs or vector databases, and essentially solves the problem of "where to store information". The real memory capability also needs to solve the problems of how information is understood, formed, updated and called.

For example, when a user instruction enters the system, the first step is to judge "what is worth remembering": what information should be recorded, in what form it should be recorded, where it should be stored, and when it should be updated, deleted and maintained; when users enter different scenarios, which historical information should be dynamically called; after further understanding the user's intention, the system also needs to accurately find the truly relevant memory from a large amount of historical information, and combine this information with the reasoning process of the foundation model, so as to finally improve the model's reasoning ability and user experience.

"It cannot be solved by simply adding a database or calling an off-the-shelf model, but is a set of system capabilities that require long-term training and engineering accumulation," said Hao Jianye. In his view, memory itself is becoming a relatively independent model capability, which is not completely the same as the general computing and reasoning capability of the foundation model itself.

Most of the earlier memory companies in the market adopted database-based solutions, which completed the storage, retrieval and management of information through infrastructure such as vector databases. It is more like adding a "memory bank" outside the model, with an external system responsible for storing and calling historical information. However, MemoraX has chosen a different "model endogenous" route from the very beginning: to make memory capability formed more through model learning, rather than relying solely on external databases for management.

This direction has been practiced by some large model companies abroad. For example, Google's Hope model attempts to design multiple nested learning loops and memory of different time scales for the model, so that the model has the ability of continuous learning and continuous evolution; Meta further attempts to embed a trainable ultra-large parameter memory layer into the Transformer architecture to enhance the memory capability of the model itself.

However, when memory begins to become an independent model capability, another problem arises: what criteria should be used to judge whether a model's memory capability is good or not?

The LoCoMo dataset is one of the earlier large-scale benchmarks specially designed to evaluate ultra-long-term conversational memory. Its conversations can span up to 35 sessions and 300 turns, which has greatly increased the memory span compared with previous conversation data that usually only covers a few sessions. However, as the evaluation goes deeper, problems have gradually emerged. Since part of the data is generated by the model, the test samples themselves may have problems such as inconsistent character relationships, contradictory facts, and chaotic timelines. If the "ruler" used to measure the model's capability is not accurate enough, the final scores can hardly truly reflect the model's memory capability.

Therefore, MemoraX AI has started to do two things at the same time: on the one hand, it corrects and optimizes the existing Benchmark, and on the other hand, it tries to build memory evaluation data that is closer to the real world.

To this end, MemoraX AI does not fully rely on artificially constructed standardized data, but tries to extract dialogue content from real scripts such as TV series and soap operas. This kind of data is closer to the interaction in the real world: complex character relationships, long-lasting dialogues, and continuously accumulated information, where the same character may mention things that happened a long time ago. For memory models, this scenario is closer to real user scenarios than simply remembering an isolated piece of information.

At present, MemoraX AI's core memory system has achieved a score of 82.65 on the LoCoMo-Refined evaluation benchmark jointly released by Nanjing University and Shanghai AI LAB, ranking SOTA, 30% higher than the second-ranked memory framework.

In addition, MemoraX AI has jointly built and open-sourced ScriptMem, an Agent Memory evaluation benchmark based on real scripts, with the University of Oxford. On this benchmark, MemoraX AI has achieved an accuracy rate of 60.3%, 40% higher than the second place, trying to further promote large model memory from laboratory tests to more real life scenarios. In the AI Coding direction, MemoraX AI has also jointly built and open-sourced SWE-context-bench, an Agent Memory evaluation benchmark for the Coding field, with the University of Oxford. On this benchmark, MemoraX AI has achieved a task completion rate of 45%, 50% higher than the second place, further verifying the memory capability in real software development tasks.

In addition to Benchmark, the company is also conducting continuous research in directions such as post-training of large model reinforcement learning and Test-Time Scaling. Focusing on these core technical directions, MemoraX AI has carried out collaborative innovation with many domestic universities, and 10 papers have been selected for ICML 2026 so far. In the basic theory of intelligent decision-making, the founding team has also won the first prize of Natural Science Award of China Electronics Society.

All these efforts point to the goal of turning "memory" from an abstract technical concept into a model capability that can be truly measured, verified, and finally applied in real scenarios.

In the Second Half of AI Development, Memory May Become the Core Competition Track

Olivia Moore, a partner at a16z, mentioned in an interview in March this year that the memory function of AI products will become a core advantage in the future. As the model remembers users' habits and contexts, products will no longer require lengthy onboarding processes, but provide highly personalized experiences from the very first interaction.

Hao Jianye judges that as public data is continuously digested by large models, the important direction of AI in the next stage is likely to further sink to every user, moving from general intelligence to personalized intelligence. At this stage, the model should not only "know the world", but also "know you" — understand users' habits, preferences, historical behaviors and long-term goals, and continuously update itself through continuous interactions. Therefore, memory is no longer a nice-to-have feature, but the infrastructure for AI to realize personalization.

Around this direction, MemoraX AI currently divides memory capability into three categories according to data form: text, code and multi-modal, and builds corresponding memory models for different data types, finally integrating the three capabilities into a general and complete memory solution.

In the AI Coding scenario, MemoraX AI has launched a memory product for Coding Agents, providing long-term project memory for programming Agents such as Codex and Claude Code, enabling them to continuously understand project contexts. In complex and long-term software development tasks, the Agent does not need to repeatedly ask users to explain previous code and task backgrounds, thus reducing context gaps and repeated communication, improving task execution effect while reducing Token consumption.

In the interactive entertainment scenario, memory capability corresponds to another demand. Even if traditional AI characters can carry out natural conversations, they often find it difficult to maintain the consistency of character relationships, story progress and world state after long-term interactions. MemoraX AI's memory engine can support thousands of consecutive rounds of interactions, dynamically precipitating and recalling key information during the interaction process, so that the characters can continuously remember what happened before. The interaction between users and AI can further evolve into a long-term memory-driven interactive experience.

In a wider range of industrial scenarios, MemoraX AI has begun to embed long-term memory capabilities into real businesses. At present, the company has carried out cooperation with customers in the fields of medical health, smart home, smart wearable, embodied intelligence and other fields. For example, it can continuously understand users' long-term health status in chronic disease management, precipitate family members' living habits and device preferences in whole-house smart scenarios, and support cross-cycle personalized interaction on smart terminals.

In terms of industrial cooperation, MemoraX AI has officially signed a strategic cooperation agreement with Huawei Cloud, and has become the only enterprise focused on the long-term memory track selected into Huawei Cloud's large model strategic cooperation vendors.

It is reported that with the completion of this round of financing, MemoraX AI will continue to tackle the underlying core technologies of large model long-term memory, further consolidating the company's technical capabilities in the AI memory field; at the same time, it will accelerate multi-scenario implementation, promote the standardized R&D and large-scale commercialization of memory modules for B-end industries and C-end interactive products.

This competition around "memory" has just begun.

This article is from WeChat official account "ChinaVenture", Author: Wei Xianghui, Editor: Liu Yanqiu, published with authorization from 36Kr.