JMIR AI

AI-powered mental health tools are increasingly deployed to support users across multiple sessions, yet the field lacks a principled framework for how memory in these systems should be structured and applied. In most current implementations, memory functions primarily as a personalization mechanism, optimizing for conversational continuity and user engagement without distinguishing between types of information that have fundamentally different clinical relevance. We propose a framework organizing memory in AI-powered mental health systems into 4 functionally distinct types. Episodic memory captures discrete, time-bound experiences tied to specific events and context. Pattern memory, adapted from the concept of procedural memory in cognitive psychology, tracks recurring patterns in cognition, emotion, and behavior across sessions. Semantic memory captures stable, personally relevant background context about the user. State-responsive memory represents the user’s current emotional and psychological condition in real time, taking priority over the other 3 types when acute distress or risk is signaled. Each type corresponds to a distinct therapeutically relevant function, and together they are designed to support the kind of cumulative, longitudinal understanding that effective mental health care requires. We term this framework “therapeutically informed memory,” drawing on established memory systems research and applying it to the clinical requirements of AI-powered mental health support. The aim of this viewpoint paper is to give AI developers, clinicians, and mental health organizations a shared vocabulary and design framework for organizing memory around therapeutic function rather than personalization alone. This paper is intended primarily for AI product and engineering teams, clinical advisors to digital mental health companies, and researchers evaluating AI-powered mental health tools. We describe the design requirements and clinical rationale for each memory type, discuss how the types interact and how priority should be assigned across them, and use Yuna, an AI-powered digital mental health intervention developed with clinical input, as an illustrative example of how this framework can be applied in practice. We conclude with design implications for the field and identify open questions regarding memory quality metrics, outcome validation, and the ethical dimensions of persistent memory as priorities for future research.

AI chatbots are increasingly used for emotional support, often outside formal care. Emerging evidence suggests that generative systems may produce modest reductions in symptoms of depression and anxiety. Yet a change in symptoms alone does not clarify the nature of the process involved. This viewpoint argues that chatbot interactions may primarily operate through structured mirroring and self-reflection rather than through the intersubjective engagement that underpins therapeutic transformation. We propose an implementation-focused distinction between tools that support symptom regulation and engagement and interventions that aim at enduring psychological change, emphasizing the need to specify mechanisms, relational conditions, and appropriate care pathways. Conflating perceived support with psychological change risks redefining therapeutic standards around technological affordance rather than relational encounters. As AI chatbots become embedded in mental health ecosystems, conceptual clarity and digital emotional literacy are essential to ensure that innovation strengthens rather than replaces the relational foundations of care.

Suicidal behavior is a major public health problem worldwide. The exact etiology remains unclear, representing a complex problem involving multiple factors. Evidence indicates that around 50% to 80% of people who die by suicide have had contact with the health care system in the year prior to their death.

Effective diabetes management requires continuous interpretation of glycemic trends, personalized dietary guidance, and sustained patient education. Although large language models (LLMs) are increasingly being explored for health-related applications, existing general-purpose and biomedical models often struggle with diabetes-specific reasoning and instruction-following, limiting their reliability for domain-focused tasks such as clinical question answering and dietary recommendation tasks.

Motivational interviewing (MI) is widely used in preventive interventions, yet coding MI techniques and monitoring intervention adherence remain resource-intensive due to the reliance on manual transcription and expert review. Large language models (LLMs) offer a promising approach to automate these tasks, but their agreement with human coders in the context of prevention interventions has not been established.

Fairness evaluation is essential for trustworthy clinical risk prediction. However, existing fairness-oriented discrimination metrics either ignore cross-group comparisons or rely on exhaustive pairwise evaluations, making them difficult to interpret and impractical for model selection.

Palliative care is increasingly recognized as essential for an aging population and rising life-limiting illnesses. Machine learning (ML) has been widely applied in this field, primarily for prognostication. However, recent literature suggests broader applications that may enhance patient-centered care and optimize system-level processes.

Multidisciplinary tumor boards (MDTs) constitute the foundation of modern tumor therapy. Large language models (LLMs) are widely discussed for optimizing their recommendations.

Maintenance of oncology clinical practice guidelines (CPGs) is increasingly challenged by the rapid growth of trial data and therapeutic complexity. While large language models (LLMs) have shown promise in information retrieval, their utility in the rigorous, end-to-end workflow of guideline maintenance remains underexplored.

Incidental detection of abdominal aortic aneurysms (AAAs) has increased with widespread cross-sectional imaging, while traditional surveillance remains fragmented and clinician-dependent. Real-world descriptions of centralized digital surveillance programs combining structured electronic health record (EHR) queries with natural language processing (NLP) of radiology reports remain limited.

Medication adherence remains a significant concern in both clinical practice and public health. Nonadherence to prescribed medication regimens is associated with poorer health outcomes, higher rates of hospitalization, and increased financial burdens on health care systems. Despite its critical role in ensuring treatment efficacy, adherence assessment is still largely dependent on patient self-reports, pharmacy refill records, or caregiver observations—methods that are often subjective, inconsistent, and unreliable. Therefore, there is an urgent need for objective, automated solutions to accurately monitor medication, particularly the information on the pills taken and intake behavior. AI techniques, especially computer vision, are promising solutions; however, their application in assessing medication adherence has several key pitfalls in maintaining robustness under environmental variation that limit their reliability and real-world applicability.

AI research increasingly depends on energy-intensive computation, yet energy use, greenhouse gas emissions, hardware life cycle burdens, and water consumption are rarely reported in a standardized way. This lack of reproducible environmental accounting limits comparisons across studies and obscures the trade-offs among model performance, infrastructure choices, carbon intensity, and cooling water demand.