Auri AI companion uses CTEM to feel emotions over time

A team of researchers (Feier Qin, Xiao Li, Yi Zheng, et al.) from an undisclosed institution has introduced Cross-Temporal Emotional Modeling (CTEM), a framework designed to make conversational agents feel like long-term companions rather than episodic task bots. The key insight is that current agents fail to maintain a coherent emotional history: past behaviors rarely affect current emotions, and emotions seldom shape future actions. CTEM closes this loop by linking long-term behavioral history to moment-to-moment emotional expression, creating a state that updates with each interaction and user feedback.

To test CTEM, the team built Auri, an agent deployed on an instant-messaging platform. In a 21-day in-the-wild study, users interacted naturally with Auri and rated it significantly higher on perceived naturalness, coherence, and emotional harmony compared to standard episodic agents. The work was published in CHI '26 proceedings and is available on arXiv. This could pave the way toward AI companions that feel more authentic, remembering your jokes, grievances, and growth over time.