Cross-Temporal Emotion Modeling for Companionable AI Agents

Introducing a novel framework known as Cross-Temporal Emotion Modeling (CTEM), which seeks to enhance the naturalness and companionship of conversational AI agents by connecting long-term behavioral patterns with real-time emotional expressions. Existing agents struggle with prolonged interactions as their behaviors seldom affect emotional states, and emotions rarely influence future actions. CTEM establishes a feedback loop where prior experiences inform a developing emotional state, impacting immediate exchanges, while user responses adjust both memory and emotion. This framework is exemplified by Auri, a virtual agent crafted for enduring companionship. The findings are available on arXiv (2605.15812).