cs.AIApr 13, 2026
SaveA Framework for Longitudinal Health AI Agents
Organizations: Columbia University, Biomedical Informatics, New York, New York USA · Columbia University, Computer Science, New York, New York USA
Abstract
Although artificial intelligence (AI) agents are increasingly proposed to support potentially longitudinal health tasks, such as symptom management, behavior change, and patient support, most current implementations fall short of facilitating user intent and fostering accountability. This contrasts with prior work on supporting longitudinal needs, both within and beyond clinical settings, where follow-up, coherent reasoning, and sustained alignment with individuals' goals are critical for both effectiveness and safety. In this paper, we draw on established clinical and personal health informatics frameworks to define what it would mean to orchestrate longitudinal health interactions with AI agents. We propose a multi-layer framework and corresponding agent architecture that operationalizes adaptation, coherence, continuity, and agency across repeated interactions. Through representative use cases, we demonstrate how longitudinal agents can maintain meaningful engagement, adapt to evolving goals, and support safe, personalized decision-making over time. Our findings underscore both the promise and the complexity of designing systems capable of supporting health trajectories beyond isolated interactions, and we offer guidance for future research and development in multi-session, user-centered health AI.
Figures & tables
Figure 1: A four-part framework integrating coherence , continuity , adaptation , and agency to support sustained health engagement over time. Adaptation refines knowledge through responsiveness , personalization , and reflexivity , while coherence builds stable interpretive foundations ( history , organization , relationship , persistence ). Continuity sustains long-term goals via follow-up , alignment , and accountability , and agency calibrates control through negotiation , transparency , emancipation , and proactivity . The arrows illustrate possible pathways of influence among layers (e.g., adaptation informing coherence by refining what remains stable versus revisable knowledge). However, in practice, these capacities are likely deeply interwoven, interacting bidirectionally and recursively across encounters rather than operating in a fixed or linear sequence.
| Dimension | Description | Illustrative example | Potential success check |
|---|---|---|---|
| History [ 19 , 65 ] | Store past interactions alongside agent-generated hypotheses and reasoning chains linking events to outcomes. | The agent links a recent symptom or behavior to previously recorded patterns and past interventions to explain likely causes. | Whether prior interactions and reasoning are consistently and accurately incorporated into subsequent responses across sessions. |
| Organization [ 64 ] | Encode care entities, roles, and relationships explicitly in structured representations rather than implicit context embeddings. | The agent surfaces and takes into account roles (e.g., user, clinician, caregiver) and their responsibilities within an ongoing care plan. | Whether roles, relationships, and responsibilities can be reliably identified and retrieved without contradiction or loss of structure. |
| Relationship [ 52 , 19 ] | Maintain identifiable, stable interaction patterns to ensure continuity and predictable agent behavior over sessions. | The agent maintains a consistent interaction style and references prior discussions to reinforce continuity across sessions. | Whether interaction patterns (e.g., tone, references, framing) remain stable and predictable across sessions. |
| Persistence [ 17 , 69 , 64 ] | Retain core facts, commitments, and explanatory hypotheses across updates and changing circumstances. | The agent retains key facts, plans, or prior conclusions across sessions and minor user context changes. | Whether previously established information remain consistent over time, except when explicitly revised. |
Table 1: Operationalization and evaluation of the Coherence layer.
| Dimension | Description | Illustrative example | Potential success check |
|---|---|---|---|
| Follow-up [ 11 , 12 ] | Track unresolved concerns, plans, or decisions across sessions and proactively resurface them based on prior context and risk signals rather than generic prompts. | The agent revisits a previously discussed concern, plan, or decision at a later time based on its relevance or status. | Whether unresolved concerns, plans, or decisions are systematically revisited over time based on prior context without re-introduction. |
| Alignment [ 31 , 53 , 38 ] | Maintain linkage between short-term actions and evolving long-term health objectives, adjusting guidance as goals shift. | The agent connects immediate recommendations or behaviors to a user’s broader, evolving goals over time. | Whether short-term recommendations remain explicitly linked to evolving long-term goals, particularly during transitions or goal changes. |
| Accountability [ 50 , 45 , 23 , 33 ] | Monitor progress over time, acknowledging partial gains, recalibration, and non-linear outcomes instead of only completed tasks. | The agent tracks incremental progress and setbacks and incorporates them into ongoing guidance rather than resetting progress. | Whether progress, including partial gains and setbacks, is tracked and incorporated into subsequent guidance rather than reset or ignored. |
Table 2: Operationalization and evaluation of the Continuity layer.
| Dimension | Description | Illustrative example | Potential success check |
|---|---|---|---|
| Responsiveness [ 66 , 59 , 45 , 40 ] | Continuously monitor user-provided signals and surface potential external changes, adjusting reasoning and interaction logic while filtering transient noise. | The agent adjusts its recommendations when it detects meaningful changes in user state, behavior, or context. | Whether meaningful changes in user state or context result in timely and appropriate updates to recommendations, without overreacting to transient variation. |
| Personalization [ 1 , 42 ] | Tailor strategies and intervention policies dynamically to individual goals, behaviors, context, and longitudinal trajectories. | The agent tailors suggestions based on the user’s preferences, routines, and prior responses to interventions. | Whether guidance reflects individualized patterns (e.g., preferences, routines, prior responses) and adapts as these patterns evolve over time. |
| Reflexivity [ 38 , 37 ] | Re-evaluate assumptions, inferred preferences, and decision rules, updating internal models when longitudinal evidence shows misalignment with user states or objectives. | The agent revises earlier assumptions or inferred preferences when new patterns indicate they are no longer accurate. | Whether prior assumptions or inferred preferences are revised when evidence indicates misalignment, and whether such revisions are reflected in future behavior. |
Table 3: Operationalization and evaluation of the Adaptation layer.
| Dimension | Description | Illustrative example | Potential success check |
|---|---|---|---|
| Negotiation [ 2 , 58 ] | Make agent initiative configurable and contestable, allowing users to question, override, or adjust guidance over time. | The agent allows users to modify or override its recommendations and adapts future guidance accordingly. | Whether users can modify or override contest system recommendations, and whether these adjustments influence future interactions. |
| Transparency [ 2 ] | Expose reasoning, uncertainty, and decision logic in interpretable ways so users can calibrate trust and reliance. | The agent explains the reasoning behind its recommendations, including relevant inputs and uncertainty. | Whether the system’s reasoning, inputs, and uncertainty are exposed in ways that allow users to understand and evaluate its recommendations. |
| Emancipation [ 31 , 38 , 33 ] | Support user independence by providing explanations, hints, and decision templates, gradually reducing intervention as competence increases. | The agent gradually reduces guidance as the user demonstrates increased confidence and independent decision-making. | Whether user reliance on the agent decreases appropriately over time (e.g., reduced prompts) while maintaining or improving outcomes. |
| Proactivity [ 71 , 21 ] | Initiate actions or interventions appropriately while remaining aligned with user preferences and safeguards. | The agent initiates a check-in or suggestion when it detects potential risk or missed follow-up. | Whether the agent initiates timely and contextually appropriate actions aligned with user goals and preferences, without excessive or misaligned intervention. |
Table 4: Operationalization and evaluation of the Agency layer.
Figure 2: A potential timeline of a longitudinal health agent supporting an individual with endometriosis. The agent maintains coherence by integrating past symptoms and interventions, ensures continuity by actively tracking strategies and follow-ups, enables adaptation by recalibrating guidance as symptoms and context change, and fosters agency by scaffolding patient reflection and self-management.
| Longitudinal Health Agent Design | Open questions |
|---|---|
| Persistent memory and structured context ( coherence ) | • How can agents represent and update structured longitudinal knowledge to maintain coherence across sessions while allowing safe revision? • What are effective ways to encode relationships, roles, and hypotheses for evolving health contexts? |
| Tracking evolving goals ( continuity ) | • How can agents dynamically capture partial progress and non-linear trajectories? • What methods best link short-term tasks to long-term objectives and maintain alignment with shifting user priorities? |
| Reflexive adaptation ( adaptation ) | • How can agents detect misalignment between past inferences and current internal and external evidence, and update strategies without disrupting continuity? • How should second-order adaptation be evaluated longitudinally? |
| Longitudinal agency calibration ( agency ) | • How can agents negotiate initiative over time, balancing proactive support and user autonomy? • What metrics can capture user empowerment, trust, and decision-making confidence across extended use? |
| Evaluation of longitudinal outcomes | • What study designs, metrics, and qualitative approaches can capture performance across coherence, continuity, adaptation, and agency over time? • How can longitudinal impact on user trajectories be assessed systematically? |
Table 5: Potential research agenda for longitudinal health agents. The table highlights open questions for each key capability of the framework.