A Self-Evolving Agent for Longitudinal Personal Hea... | AI Research

A Self-Evolving Agent for Longitudinal Personal Health Management
Most current health AI systems treat every user interaction as a fresh start, failing to remember a person’s unique history, changing routines, or evolving health risks. This paper introduces HealthClaw, an open-source agent architecture designed to provide continuous, personalized health support. By moving beyond isolated queries, HealthClaw creates a "self-evolving" memory that updates over time, allowing the system to learn from past encounters while maintaining strict boundaries for privacy and safety.

A Smarter Way to Remember

HealthClaw functions as a closed loop that processes information in four stages: perception, reasoning, action, and post-episode induction. The key innovation is the induction step, which occurs after each interaction. Instead of simply saving the entire conversation—which can be cluttered and privacy-risky—the system intelligently decides what to keep. It categorizes information into different "memory layers," such as durable profile facts (like allergies), reusable procedures (like meal-planning habits), and temporary episodic traces. This allows the agent to build a personalized history that is both relevant and manageable.

Balancing Recall and Privacy

A major challenge in health AI is the trade-off between having enough information to be helpful and protecting sensitive data. The researchers compared HealthClaw against two baselines: "current-only" prompting (which has no memory) and "full-history" prompting (which provides the entire past record). HealthClaw significantly outperformed the current-only model in accuracy and proved more effective at maintaining privacy than the full-history approach. By selectively retrieving only the necessary information, HealthClaw reduced the amount of context data exposed by 71.7% compared to full-history methods, while also reducing the frequency of unsafe disclosures.

Performance Across Biomedical Tasks

Beyond longitudinal support, the researchers tested HealthClaw on nine diverse biomedical tasks, including analyzing medical imaging, genomic data, and physiological signals. The agent consistently improved performance across these tasks, achieving a mean absolute gain of 27.0 percentage points in primary metrics. These results suggest that the agent’s ability to route and integrate heterogeneous health evidence is a powerful tool for clinical decision support, helping the system adapt to different types of medical data as they arise.

Important Considerations

While these results are promising, the researchers emphasize that HealthClaw has only been evaluated through offline, simulated benchmarks. The study used a synthetic year-long trajectory and automated grading, which cannot fully replicate the complexities of real-world clinical environments, such as human judgment, unexpected behavioral shifts, or the nuances of doctor-patient interactions. As with any medical AI, the authors note that clinical effectiveness must be proven through prospective, real-world studies before the system can be considered for actual medical practice.