Emotion-Attended Stateful Memory Architecture Enhances LLM Personalization

A novel architecture known as Emotion-Attended Stateful Memory (EASM) has been introduced to tackle the inherent stateless nature of existing language model systems during interactions. This method creates a personalized conversational context by leveraging long-term history, emotional cues, and inferred intentions in real-time. In a controlled A/B experiment involving thirty unscripted dialogues across six distinct emotional categories, the memory-enhanced condition consistently surpassed the stateless baseline. Notable improvements were recorded in memory grounding (95% enhancement), plan clarity (57%), and emotional validation (34%). This research has been made available on arXiv with the identifier 2605.14833.