HealthClaw: A Self-Evolving Agent for Longitudinal Personal Health Management

Introduction

Personal health management is rarely a one-off question. Diet, exercise, medication habits, sleep, measurements, and risk factors evolve over repeated encounters. Yet many health AI systems still respond as if every request starts from zero. A new arXiv paper proposes HealthClaw, an open-source agent architecture that aims to make health support more continuous by maintaining governed, self-evolving memory over time.

Key points

• From isolated queries to longitudinal support: HealthClaw is designed for scenarios where a person’s routines, preferences, health measurements, and risks change across months of interaction.
• Separation of knowledge and private memory: The architecture keeps shared safety rules and medical knowledge distinct from private longitudinal memory. That memory contains profile facts, reusable procedures, and episodic traces.
• Post-episode induction: After each episode, the system determines whether new information should update the user profile, revise a reusable procedure, remain as an episodic record, or be excluded.
• Better performance with less prompt exposure: Across 900 longitudinal support probes, answer accuracy rose from 0.2% with current-query prompting to 45.7% with HealthClaw. Compared with full-history prompting, prompt-side context exposure was reduced by 71.7%.
• Improved privacy handling: In 100 privacy probes, HealthClaw produced higher-quality privacy-aware answers and fewer unsafe disclosures than both baseline methods.
• Gains on biomedical tasks: Across nine biomedical tasks with 200 cases each, the mean absolute improvement in the task-specific primary metric was 27.0 percentage points. Seven gains remained significant after false-discovery-rate correction.

Why it matters

The paper addresses a central tension in health AI: users want systems to remember relevant context, but sensitive histories should not be indiscriminately injected into every prompt. HealthClaw’s contribution is not simply adding more memory. It proposes a governance layer that decides what should become stable profile knowledge, what should become a reusable procedure, what should stay as a past episode, and what should not be stored.

This approach is especially relevant for high-stakes domains where personalization and privacy must coexist. In health management, an agent that remembers exercise limitations, medication routines, or preferred communication styles could offer more relevant support. But the same memory, if poorly governed, could create privacy and safety risks. HealthClaw’s design attempts to reduce that risk by limiting prompt-side exposure while retaining useful continuity.

Still, the results should be read carefully. The evaluation relies on a synthetic year-long benchmark and offline biomedical tasks. The reported 45.7% accuracy is a major jump over current-query prompting, but it is not evidence that the system is ready for clinical deployment. The authors explicitly state that prospective evaluation is needed to assess clinical effectiveness.

HealthClaw is therefore best understood as an architecture-level step toward long-term personal health agents. Its broader lesson may extend beyond medicine: memory in AI systems should be selective, structured, auditable, and governed—not just longer.

Source: arXiv