Tencent’s PCG team has open-sourced T-Mem, a long-term memory system whose paper has been accepted to the main conference of EMNLP 2026. The project focuses on a specific weakness in conventional memory systems: retrieval breaks down when a new query shares few or no keywords with earlier stored information. T-Mem addresses that by recording not just a fact, but also the kinds of future situations where that fact may become useful.
The example cited in the release is a user saying a colleague is allergic to seafood, then later asking where the team should go for a group meal. A standard memory system may fail to connect the two because the wording barely overlaps. T-Mem is designed to link "seafood allergy" with the scenario of choosing a restaurant for a gathering, allowing the earlier memory to be recalled later. Evaluation results said ordinary memory systems saw performance drop sharply in low-keyword-overlap settings, while T-Mem was less affected. The report also said results declined noticeably after removing Horizon Trigger, the component used to predict future scenarios, identifying it as central to T-Mem’s associative recall.
Tencent’s PCG team has open-sourced T-Mem, a long-term memory system, and its paper has been accepted to the main conference of EMNLP 2026.
T-Mem is built to store more than a single piece of information. When it records a fact, it also notes the kinds of situations where that fact may become useful later.
The example given is a user saying that a colleague is allergic to seafood, then later asking where the team should go for a group dinner. Those two statements share very few keywords, so a conventional memory system may fail to retrieve the earlier detail. T-Mem is designed to connect "seafood allergy" with the scenario of choosing a restaurant for a group meal, making that earlier memory retrievable in a related context.
According to the evaluation described in the release, standard memory systems see performance drop sharply when a new question has little keyword overlap with past memories. T-Mem was less affected under the same condition.
The report also said that removing Horizon Trigger, the module responsible for predicting future scenarios, led to a clear decline in T-Mem’s results. That component was identified as the key part behind its associative recall capability.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
