MemoraX AI completes a seed round financing of tens of millions of US dollars, defining a new paradigm of "endogenous memory" for large models.

Recently, Shenzhen MemoraX AI Technology Co., Ltd. (hereinafter referred to as "MemoraX AI" or "MemoraX") announced the completion of a seed - round financing of tens of millions of US dollars. This round was co - led by the L2F Light Source Entrepreneurs Fund and Zhongding Capital, with well - known investors and industry players participating together. Light Source Capital served as the exclusive financial advisor.

The funds from this round of financing will be mainly used for the algorithm iteration and engineering implementation of the core technology Agentic RL and the product development of the endogenous memory module. It is expected that within the next 12 months, MemoraX AI will launch the first batch of standardized memory products for B - end enterprise knowledge management and C - end personalized interaction, achieving an exponential improvement in memory training efficiency and accuracy. This financing marks that the large - model personalized memory track has officially moved from the shallow exploration of "external retrieval" to the in - depth transformation of "model endogeny", injecting core impetus into solving the key bottleneck of the large - model "amnesia".

A "battle - tested" dream team: Top - ten scientists from top conferences × Industrial operators with tens of billions of business experience

MemoraX AI was founded in March 2026 and is a promoter of technological innovation in the large - model personalized memory track. Through the independently developed Agentic RL (Agent Reinforcement Learning) core technology architecture, the company has overcome the industry pain points of large - models, such as fragmented long - term memory, static memory, and difficulty in cross - scenario migration.

The founder, Haojianye, is an elite professor at Tianjin University, a national outstanding young scientist, and one of the top 2% scientists globally. In the past two years, his research output at the three top conferences ICML, NeurIPS, and ICLR ranked among the top 10 globally, and his Google Scholar citations exceeded 15,000 times. More importantly, he is not a traditional academic entrepreneur. He has successively served as the director of the Decision - Inference Laboratory at Huawei, the director of the Large - Model Algorithm Laboratory at Huawei, and the technology president of the Huawei Medical Legion. He is the chief expert in the field of decision - making intelligence at Huawei and led the incubation of several major industrial projects from scratch, creating economic benefits of tens of billions of yuan.

The core members of the team mainly come from two systems. On the one hand, they are core technology leaders from leading companies such as Huawei, Alibaba, and Tencent, who have been responsible for the R & D of large - models with a scale of hundreds of billions and the implementation of Agents, and have complete experience from underlying technology to commercial deployment. On the other hand, they are the most representative top research forces in the field of domestic reinforcement learning, with both algorithm innovation and industrial implementation experience. They have long - term in - depth cooperation with industrial teams such as Alibaba, Tencent, NetEase, and ByteDance. Their cooperation results have been widely implemented in industrial - level scenarios such as autonomous driving, game AI, recommendation search, and embodied intelligence, creating economic benefits of billions of yuan. Pure academic teams are strong in exploration but weak in implementation, while pure industrial teams are good at product development but difficult to build barriers. MemoraX AI has the genes of both, forming a full - link closed - loop ability from theoretical innovation to industrial implementation.

Not an external attachment but an endogenous solution: A "paradigm revolution" in large - model memory

The long - term memory track is not an uncharted area. However, most of the existing solutions remain at a shallow level of exploration, essentially still being "static information retrieval". It's like giving AI a notebook; it can only look through and search but cannot truly remember you.

MemoraX AI has chosen a more difficult but more fundamental path: internalizing memory ability into the model. This means that instead of attaching an external memory bank to AI, they make memory an endogenous ability of the model. AI is no longer just an information retriever but can actively learn, refine, precipitate, and dynamically update memory like a human being.

This is a revolutionary breakthrough from external memory attachment to model endogeny. The technological features of MemoraX AI are reflected in three dimensions:

First, the continuous evolution ability of memory. Memory is no longer a static collection of stored information but a process of continuous understanding, updating, and reorganization in interaction, thus achieving accumulation over time, correction with feedback, and migration across scenarios.

Second, the accurate recall ability of memory. It solves the pain points of "fragmented memory and inaccurate retrieval" in traditional solutions. It achieved the strongest memory performance on the latest text - based memory test set LoCoMo - Refined (significantly leading the second - place by 30%), and the model training efficiency was improved by 400 times.

Third, the generalization and reuse ability of memory. It can achieve cross - scenario memory reuse and can be quickly adapted to scenarios such as next - generation intelligent interaction, enterprise - level knowledge management, personalized digital companionship, AI Coding, and new - generation intelligent terminals.

Agentic RL: The key technology for building the next - generation memory system

The core ability breakthrough of the next - generation memory system lies not only in storing user information but also in achieving accurate perception of context, efficient compression and purification of information, dynamic recall and flexible assembly of memory, and adaptive evolution towards users' long - term preferences in the process of dynamic interaction. Agentic RL is precisely the key technological path to achieve the above - mentioned high - order memory capabilities. In this direction, the MemoraX AI team has fully proven itself. A series of cutting - edge Agentic RL technologies proposed by the team have been implemented on a large scale in multiple industrial scenarios such as chip design automation, industrial solvers, autonomous driving, and entertainment intelligence:

(1) In the field of chip design automation (EDA), the intelligent decision - making technology based on reinforcement learning built by the team ranked first in the international authoritative list EPFL in the field of chip logic synthesis for two consecutive years. It has been implemented in important chip design links such as chip logic synthesis and physical layout and applied in the design of dozens of domestic chips, significantly improving chip design efficiency and PPA comprehensive performance.

(2) In the field of industrial solvers, the optimization algorithm for solving problems based on reinforcement learning proposed by the team has significantly improved the solving efficiency of solvers in large - scale solving problems. It surpassed the solving performance of the international commercial solver Gurobi for the first time on the global authoritative evaluation list, won the championship, and received the highest award SAIL at the World Artificial Intelligence Conference.

(3) In the field of autonomous driving, the intelligent driving algorithm based on reinforcement learning developed by the team has been deployed in hundreds of thousands of autonomous vehicles, significantly improving the human - like driving experience. It is the first case in the industry where reinforcement learning technology has been commercially implemented in the field of autonomous driving.

(4) In the field of game AI, the team developed the industry's first automated game testing technology based on reinforcement learning and the multi - style AI generation technology driven by reinforcement learning. These technologies have been applied in more than 10 large - scale commercial games of NetEase and won the first - class award for scientific and technological progress from the Chinese Society of Image and Graphics.

(5) In the field of advertising recommendation, the team developed the industry's first advertising bidding algorithm based on hierarchical reinforcement learning, which has been applied in the Taobao information flow scenario, helping Taobao's targeted advertising daily revenue exceed 100 million yuan.

These experiences mean that the MemoraX AI team not only understands technology but also knows how to transform technology into products and continuously create commercial value. For the memory track, a team like MemoraX AI, which has both top - level R & D and industrial implementation capabilities, is extremely scarce and has the greatest chance to achieve a core breakthrough first and finally run through the commercial closed - loop.

From "tool" to "partner", redefining the relationship between humans and AI

MemoraX AI's business layout is a two - way strategy of integrating hardware and software.

On the B - end, MemoraX AI's standardized memory modules can empower intelligent customer service, enterprise knowledge management, and professional fields such as finance, medical care, and law, breaking the pain point of repeated inquiries and achieving accurate services. On the C - end, it aims at the anchor point of personal intelligent assistants, breaking the homogenization dilemma of the current general AI assistants of "one - size - fits - all" and creating a "personalized" exclusive intelligent partner that can deeply engrave users' habits, preferences, learning rhythms, and work requirements.

MemoraX AI's vision goes far beyond commercial success.

"Memory is the soul of intelligence." In the view of Haojianye, if large - models cannot cross the gap from "storage" to "memory", they will always be just efficient search engines rather than real intelligent partners. By enabling AI to have real memory ability, it is not only promoting technological progress but also redefining the relationship between humans and AI, allowing AI to evolve from a cold "tool" to a warm, intelligent, and memory - enabled "partner".

This is a story about forgetting and remembering. In the long evolutionary history of AI, MemoraX AI wants to be the starting point for machines to truly "remember" humans.

Zheng Xuanle, the founder and CEO of Light Source Capital and the founding partner of the L2F Light Source Entrepreneurs Fund, said: "The Memory system is one of the most core infrastructure capabilities of Agents, directly determining the upper limit of Agent experience and delivery capabilities. Whether it is the Agent software track for companionship or efficiency tools, or the hardware battlefield with AI - native features as the core value proposition, building a reliable, long - term, and stable Memory ability is the most urgent bottleneck to be broken through in the current AI application layer. The essence of this problem is how to enable the model to continuously learn and self - update in dynamic interaction, which is exactly the core proposition that MemoraX AI focuses on solving. Professor Haojianye, the founder, has been deeply involved in the RL field for many years and has a strong track record in both academia and industry. We believe he is the best candidate to truly push this direction from research to product implementation. Therefore, Light Source was deeply involved in the construction and incubation of the project before MemoraX AI was officially established, and we also look forward to continuously coordinating the diversified capabilities of the Light Source product matrix to provide more solid support for the growth of MemoraX AI."

Li Hao, a partner at Light Source Capital, said: "We highly recognize MemoraX AI's new paradigm of redefining large - model memory with the Agentic RL framework, which truly internalizes memory ability into the model's native ability, achieving three major breakthroughs in continuous evolution, accurate recall, and cross - scenario generalization. It achieved SOTA on the LoCoMo - Refined test set, leading the second - place by 30%, and the training efficiency was improved by hundreds of times. We firmly believe that relying on the innovative technical framework of Teacher Haojianye and the core team, combined with the team's past experience in R & D of large - models with a scale of hundreds of billions and industrial implementation in multiple scenarios, will continuously bring new 'partner - style' memory ability breakthroughs to various fields. We look forward to joining hands with MemoraX AI to jointly promote the transformation of large - models from 'instant dialogue tools' to 'intelligent partners with long - term memory', opening a new era of AI - native memory."

Liu Mengsu, an executive director at Light Source Capital, said: "We are honored to assist MemoraX AI in completing this round of financing. Different from the common algorithmic incremental optimization in the industry, MemoraX AI is committed to the underlying paradigm innovation of the large - model memory mechanism. Its first - created 'endogenous memory' architecture fundamentally breaks through the ability bottlenecks of fragmentation and staticness in traditional external attachment solutions, promoting the transformation of AI from an 'information retriever' to a memory subject that 'actively learns and dynamically precipitates'. Memory is the cornerstone of intelligence. We firmly believe that relying on the core fulcrum of native memory ability, MemoraX AI will transform every interaction into continuous value accumulation, and jointly lead all industries into a new AI era of 'the more you use it, the more it understands you'."