Professor Hao Jianye from Tianjin University founded MemoraX AI to turn agent memory into a trainable learning problem, aiming to move AI from a model-centric era to an experience-driven one. He argues that true memory goes beyond storage and retrieval, requiring decisions about what to remember, when to update, and what to forget.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
As large model context windows expand, long-term memory for agents remains a weak point. MemoraX AI, founded by Tianjin University professor Hao Jianye, attempts to turn 'what to remember and what to forget' into an end-to-end trainable problem with learnable memory models, advocating that memory should be a neutral infrastructure layer independent of models.
Hao Jianye's academic and industrial trajectory spans reinforcement learning, multi-agent systems, and deep reinforcement learning applications in game AI, intelligent recommendation, autonomous driving, network optimization, and chip design. In his view, from his doctoral studies to founding MemoraX, the underlying problem has remained unchanged: how can an agent leverage past experiences during continuous interaction to improve future decision-making?
The emergence of large models and agents has changed the task environment and information forms, but not this essence. When agents face not structured states and rewards but natural language, code, user preferences, and cross-month project experiences, 'how to remember the past' becomes a prerequisite for 'using the past.'
Hao Jianye clearly distinguishes between context, RAG, and true Memory. Context merely flattens raw information for the model, and RAG passively retrieves via similarity search; neither addresses active understanding, filtering, and distillation of information.
True Memory must possess three core features: active filtering and extraction (identifying what is worth long-term retention), dynamic management and updating (continuously integrating and correcting with new information), and intelligent invocation and experience reuse (proactively recalling at the right time and distilling into transferable experience).
This article's information primarily comes from DeepTech's exclusive interview with Hao Jianye, representing the founder's own account. Some data (such as test scores and costs) are from MemoraX's internal test results, not independently verified by third parties, and should be regarded as source_claim.
Primary report
Primary source