memU

MemU is a memory-system startup founded in 2025 by Chen Hong, a former ByteDance Seed researcher who led memory algorithms for Doubao and Maoxiang (the company's AI companion products) . The company builds what it calls "Memory as File System"—a user memory management layer for long-term emotional companionship applications, using a file-system architecture rather than vector databases and adding a memory-administrator agent for more accurate, efficient recall at lower cost .

Chen, born in 1993, holds a PhD in NLP from the University of Tokyo and worked on TikTok's customer-service robot before joining Seed . He has said he chose infrastructure over building end-user products because he recognized his own limitations on product design—"I kept trying to shove in the coolest tech without knowing what users actually needed"—and instead wanted to enable other teams to build better companion experiences . MemU's open-source implementation gained 1,400 GitHub stars within two weeks of release and scored 92.09% average accuracy on the Locomo reasoning benchmark ; the team maintains a commercial version one to two releases ahead of the open-source edition .

The six-person team targets B2B customers in emotionally intensive verticals—AI companions, role-play, original character (OC) communities, education, and companion hardware—rather than tool-oriented agents like Manus or Genspark, where user memory is a lower priority . As of August 2025, the product had just launched and Chen was in early talks with enterprise clients, with funding details undisclosed .