A new arXiv paper defines a multi-layer architecture for AI agents supporting longitudinal health tasks including symptom management, behavior change and patient support. (https://arxiv.org/abs/2604.12019)

Current implementations fall short on user intent, follow-up and goal alignment according to established clinical and personal health informatics frameworks cited in the work (Lin et al., arXiv:2604.12019). The architecture operationalizes four elements: adaptation to evolving user goals, coherent reasoning across sessions, continuity of context, and support for user agency, illustrated via use cases showing maintained engagement and safer decision-making over time. This aligns with earlier personal informatics models (Li et al., CHI 2010) and agent memory mechanisms surveyed in Zhong et al., arXiv:2404.13501.

Mainstream reporting on health AI has reduced such systems to generic chatbots, missing the longitudinal pattern of persistent companions that synthesize multi-year user trajectories rather than resetting per interaction. The submitted framework corrects this by drawing directly on chronic care models requiring repeated alignment, yet leaves implicit the integration barriers with electronic health records noted in Topol, NEJM AI 2023.

Synthesis of the primary arXiv submission with the cited informatics literature and NEJM AI review reveals an emerging class of relational health agents whose safety and effectiveness scale with interaction count, shifting design focus from isolated accuracy to sustained accountability across health trajectories.