This research paper, jointly published by the Hong Kong University of Science and Technology and Nanjing University of Technology in February 2026, titled "MemFly: On the Fly Memory Optimization via Information Bottleneck" (paper number: arXiv: 2602.07885v1), has brought revolutionary breakthroughs to the memory management of large language model agents.

Computers and humans have vastly different ways of memory. When we recall the past, our brain automatically filters out irrelevant information and only retains the most important memory fragments. But nowadays, AI agents are like obsessive-compulsive disorder patients, saving every detail in its entirety. As a result, their memory becomes increasingly bloated and finding information becomes increasingly difficult.

Imagine you are organizing a huge library. Traditional AI agents would pile up every book and piece of paper intact, and over time, the entire library would become a mess. The MemFly system developed by the HKUST team is like an experienced librarian, actively organizing and categorizing similar content, merging important information in a better way, and discarding truly useless junk information.