HyT Capital Portfolio | MemoraX AI Completes Three Financing Rounds Totaling Hundreds of Millions of RMB Within Six Months

01 Investors Flock to Huawei's Former Chief Expert

The AI circle is never short of top talent, but entrepreneurs with Hao Jianye's level of excellence across academia and industry are rare.

Hao Jianye received his Ph.D. from the Chinese University of Hong Kong, under the supervision of Professor Ho-fung Leung, with research focused on game intelligence. After graduation, he worked as a postdoctoral researcher at MIT's Computer Science and Artificial Intelligence Laboratory and at the Singapore University of Technology and Design.

In 2015, Hao returned to China and joined Tianjin University. Around that time, AlphaGo burst onto the scene, sparking a new wave of interest in reinforcement learning. At that point, relevant Chinese-language communities and resources were still very limited. Hao led the establishment of the Deep Reinforcement Learning Lab at Tianjin University, making it one of the earliest research teams in China to systematically pursue this direction.

He later became a Distinguished Professor at Tianjin University, received the National Excellent Young Scientist Award, and was named among the top 2% of scientists worldwide. In the past two years, his research output at the three top conferences — ICML, NeurIPS, and ICLR — ranked in the global top 10, with over 15,000 Google Scholar citations.

What sets Hao apart from a purely academic background is that he has experienced almost every key phase of AI's journey from the lab to industry.

For years, he worked on reinforcement learning and applied deep learning technologies across a wide range of industrial deployments, covering autonomous driving, Huawei's ICT product lines, supply chains, and chip design. Before joining MemoraX AI, he served as Director of Huawei's Decision Reasoning Lab, Director of Huawei's Large Model Algorithm Lab, Technical President of Huawei's Healthcare Division, and Chief Expert in Decision Intelligence at Huawei.

During his time at Huawei, Hao led the incubation of multiple industrial projects from 0 to 1, generating tens of billions of RMB in economic value. As the industry moved into the large model era, he shifted his focus to post-training of large models, with responsibilities covering large model reinforcement learning, post-training, and industry-specific model applications.

Hao left Huawei at the end of November last year. It is understood that as early as the second half of last year, he had already begun seriously considering entrepreneurship. What made him decide was a seemingly simple yet difficult question: now that the foundation model system for large models is gradually maturing, what exactly is the core technology of the next generation of large models?

His past industry experience gave him an answer — memory.

In Hao's view, the foundation model primarily solves one-off computation and reasoning problems. But when large models truly enter various industries, interactions between users and agents will not be one-off. As interactions accumulate, users' historical behavior, preferences, task contexts, and industry knowledge all build up, and models face increasingly long contexts. Longer contexts mean higher costs and computational pressure. How to effectively save, retrieve, understand, and re-invoke large amounts of historical information at the right time becomes a new challenge beyond the foundation model itself.

Hao believes that memory is not just a nice-to-have feature — it is becoming a fundamental capability necessary for agents to achieve truly long-term interaction and continuous evolution.

02 From "Bolt-On" Memory to "Model-Intrinsic" Memory

Since the beginning of this year, model memory capabilities have attracted attention, and capital has already begun to position itself accordingly.

One example is the South Korean deep-tech company Dnotitia. In April this year, the AI memory chip-focused company closed a USD 61.2 million Series A round. Its core product, the VDPU (Vector Data Processing Unit), uses in-memory computing technology to perform matrix operations directly inside the storage array, reducing the overhead of data movement at the hardware level.

But in Hao Jianye's view, AI "memory" cannot be simplistically equated with information storage. Storage is just the underlying carrier of memory — it can be files, knowledge graphs, or vector databases — essentially solving the "where to put information" problem. True memory capability, however, also needs to address how information is understood, formed, updated, and retrieved.

For example, when a user instruction enters the system, the first step is to determine "what is worth remembering": what information should be recorded, in what form, where to store it, and when to update, delete, or maintain it. When a user enters a different scenario, the system also needs to dynamically retrieve relevant historical information. After further understanding the user's intent, it must precisely locate truly relevant memories from a large volume of historical information and integrate them with the foundation model's reasoning process, ultimately improving the model's reasoning ability and user experience.

"It's not something that can be solved by simply adding a database or calling an off-the-shelf model. It's a system-level capability that requires long-term training and engineering accumulation," said Hao Jianye. In his view, memory itself is becoming a relatively independent model capability, distinct from the general-purpose computation and reasoning capabilities of the foundation model itself.

Earlier players in the memory space mostly adopted database-style solutions, using infrastructure like vector databases for storage, retrieval, and management of information. It was more like adding a "memory bank" outside the model, with an external system handling the saving and retrieval of historical information. MemoraX, from the very beginning, chose a different path: a "model-intrinsic" approach, where memory capability is largely formed through model learning rather than managed solely through external databases.

This direction is already being explored by some major large-model companies abroad. For example, Google's Hope model attempts to design multiple nested learning loops and memory at different time scales for the model, enabling continuous learning and evolution. Meta has gone a step further, attempting to embed a trainable, ultra-large parameter memory layer into the Transformer architecture to enhance the model's inherent memory capability.

However, as memory begins to emerge as an independent model capability, another question arises: what standard should be used to judge whether a model's memory capability is good or not?

The LoCoMo dataset is one of the earliest large-scale benchmarks specifically designed for evaluating ultra-long-term conversation memory, with conversations spanning up to 35 sessions and 300 turns — a significant extension compared to earlier datasets that typically covered only a few sessions. But as evaluation deepened, issues gradually surfaced. Because some of the data was generated by models, the test samples themselves may suffer from shifting character relationships, contradictory facts, and timeline confusion. If the "ruler" used to measure model capability is itself inaccurate, then the final scores are unlikely to truly reflect the model's memory ability.

For this reason, MemoraX AI began doing two things simultaneously: n one hand, refining and optimizing existing benchmarks; on the other, attempting to construct memory evaluation data that more closely mirrors the real world.

To this end, MemoraX AI did not rely entirely on manually constructed standardized data, but began extracting dialogue content from real scripts such as TV series and soap operas. This type of data is closer to real-world interactions: complex character relationships, long conversation spans, and continuously accumulating information — the same character may mention something that happened much earlier after a long interval. For a memory model, this is much closer to real user scenarios than simply remembering an isolated piece of information.

Currently, MemoraX AI's core memory system achieved a score of 82.65 on the LoCoMo-Refined benchmark jointly released by Nanjing University and Shanghai AI Lab, ranking SOTA — a 30% improvement over the second-place memory framework.

In addition, MemoraX AI, in collaboration with the University of Oxford, constructed and open-sourced ScriptMem, an Agent Memory evaluation benchmark based on real scripts. On this benchmark, MemoraX AI achieved 60.3% accuracy, a 40% improvement over the second-place competitor, further pushing large-model memory evaluation from lab testing toward more realistic everyday scenarios. In the AI Coding direction, MemoraX AI, again in collaboration with the University of Oxford, constructed and open-sourced SWE-context-bench, an Agent Memory evaluation benchmark for the coding domain. On this benchmark, MemoraX AI achieved a 45% task resolution rate, a 50% improvement over the second-place competitor, validating memory capability in real-world software development tasks.

Beyond benchmarks, the company is also conducting ongoing research in areas such as large-model reinforcement learning post-training and Test-Time Scaling. Around these core technology directions, MemoraX AI has collaborated with multiple domestic universities on joint innovation, with a total of 10 papers accepted at ICML 2026. In terms of fundamental theory of intelligent decision-making, the founding team also received the First-Class Natural Science Award from the Chinese Institute of Electronics.

The common thread running through all this work is to turn "memory" from an abstract technical concept into a model capability that can be truly measured, truly validated, and ultimately deployed in real-world application scenarios.

03 In the Second Half of AI, the Battle May Be About Memory

In an interview this March, a16z partner Olivia Moore said that in the future, the memory function of AI products will become a core competitive advantage. As models remember user habits and context, products will no longer require lengthy onboarding — they will deliver a highly personalized experience from the very first interaction.

Hao Jianye believes that as public data continues to be absorbed by large models, the next major direction for AI may be to move further down to the individual level — shifting from general-purpose intelligence to personalized intelligence. At that stage, a model must not only "know the world," but also "know you" — understanding the user's habits, preferences, historical behavior, and long-term goals, and continuously updating them through ongoing interaction. Memory, therefore, is no longer just a nice-to-have feature, but the infrastructure that enables AI personalization.

Around this direction, MemoraX AI currently categorizes memory capabilities into three types based on data modality: text, code, and multimodal. It builds corresponding memory models for each data type, and ultimately integrates the three capabilities into a unified, comprehensive memory solution.

In the AI Coding scenario, MemoraX AI has already 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 context. In complex, long-horizon software development tasks, agents no longer need to repeatedly ask users to explain previous code and task background, reducing context gaps and redundant communication, while improving task execution and reducing token consumption.

In interactive entertainment scenarios, memory capability addresses a different need. Even when traditional AI characters can conduct natural conversations, they often struggle to maintain consistency in character relationships, story progression, and world state after long interactions. MemoraX AI's memory engine can support thousands of rounds of continuous interaction, dynamically accumulating and retrieving key information throughout the interaction, allowing characters to continuously remember what happened before. This enables interactions between users and AI to evolve further into experiences driven by long-term memory.

In broader industry scenarios, MemoraX AI has already begun embedding long-term memory capabilities into real business operations. The company is currently collaborating with clients in areas such as healthcare, smart home, smart wearables, and embodied intelligence. For example, in chronic disease management, it continuously understands the user's long-term health status; in whole-home smart systems, it accumulates household members' living habits and device preferences; and in smart terminals, it supports cross-cycle personalized interactions.

On the industry partnership front, MemoraX AI has formally signed a strategic cooperation agreement with Huawei Cloud and is the only company selected among Huawei Cloud's large-model strategic partners that focuses exclusively on the long-term memory track.

It is understood that with this financing round in place, MemoraX AI will continue to focus on core technology breakthroughs in long-term memory for large models, further strengthening its technical capabilities in the AI memory field. At the same time, the company will accelerate deployment across multiple scenarios, advancing the standardization and scaled commercialization of memory modules for both B-end industry solutions and C-end interactive products.

The competition around "memory" is just beginning.

Back Next article