Beating AI News Flash: Tencent PCG team has open-sourced a long-term memory system T-Mem, with its paper accepted to the EMNLP 2026 main conference. When it remembers something, it also records "in what scenarios this might be useful later."
For example, a user once said "my colleague is allergic to seafood," and later only asked "where should we eat for the team building." The two sentences have almost no overlapping keywords, and an ordinary memory system might not be able to find that old piece of information. T-Mem proactively links "seafood allergy" with "choosing a restaurant for a group meal," so it can retrieve it in similar scenarios later.
Evaluations show that when questions and old memories lack keyword overlap, ordinary memory systems' performance drops sharply, while T-Mem is affected much less. After removing the Horizon Trigger, which is responsible for predicting future scenarios, T-Mem's performance also declines significantly — this component is precisely the key to its "associative recall."
