The research addresses the challenge of on-device reasoning for personalized interventions within wearable systems. Affective Agent is presented as a three-layer architecture: perception, personalization, and reasoning. The perception layer utilizes physiological evidence, context, and user history. The personalization layer adapts to individual users through host-managed structured memory evolution. The reasoning layer employs a sub-billion-parameter language model to determine whether, when, and how to intervene. The system operates without requiring per-user weight updates. Evaluation was conducted using simulator-generated longitudinal scenarios, encompassing physiological variation, context, signal quality, and intervention history. Results indicated that memory-driven personalization and two-pass structured reasoning improved intervention decisions. This approach moves the decision-making process on-device, enabling closed-loop, personalized intervention on wearable hardware. The architecture is designed for inference on wearable-class hardware.
Source: https://arxiv.org/abs/2609.12322
