Will large models inevitably be efficient "search engines"? According to Hao Jianye, founder of Shenzhen Yijiayuan Technology Co., Ltd. (MemoraX AI), if the gap between "storage" and "memory" cannot be bridged, AI will never become a true intelligent partner.
Recently, this startup, which was established only a month ago, announced a $10 million seed round financing. This round was led by L2F Lightsource Entrepreneur Fund and Zhongding Capital, with participation from well-known investors and industry players. It is reported that this funding will mainly be used for the iteration of Agentic RL (agent reinforcement learning) algorithms, engineering implementation, and the product development of internal memory modules.
MemoraX AI was founded in March 2026. Its leading figure, Hao Jianye, is an elite professor at Tianjin University. He has previously served as Director of Huawei's Decision Reasoning Lab and Large Model Algorithm Lab, and holds a high academic reputation and practical experience in the field of reinforcement learning. The team consists of technical elites from big companies such as Huawei, Alibaba, and Tencent, forming a strong configuration combining academia and industry.
On the technical side, MemoraX AI aims to internalize memory capabilities through its self-developed Agentic RL technology. This solution primarily addresses common pain points of current large models, such as "fragmented memory" and "imprecise retrieval." Its core advantages are reflected in three dimensions: first, dynamic evolution, where memory can continuously update and reorganize during interactions, rather than being a static storage; second, precise recall, with performance 30% better than peers on the LoCoMo-Refined test set, and a training efficiency increase of 400 times; lastly, generalization and reuse, enabling flexible adaptation to diverse scenarios ranging from smart terminals to enterprise-level knowledge management.
Currently, these cutting-edge technologies have been initially applied in fields such as autonomous driving, chip design automation, and industrial solvers.
In terms of commercial strategy, MemoraX AI has chosen a dual approach of B-end and C-end. On the B-end, the company aims to provide standardized memory modules for industries such as healthcare, finance, and law, enhancing the depth of intelligent customer service and knowledge management. On the C-end, it is dedicated to creating more personalized digital companions. According to the plan, the first standardized memory products are expected to officially launch within a year.
Representatives from the investment side believe that a memory system is the most fundamental infrastructure for an agent (Agent), directly determining the upper limit of AI delivery capabilities. At a time when the AI application layer urgently needs to break through bottlenecks, MemoraX AI's technological path of self-updating during dynamic interactions is seen as a key attempt to break through the industry ceiling.
