A Self-Evolving Agent for Longitudinal Personal Health Management

TL;DR — HealthClaw is an open-source self-evolving agent architecture for longitudinal personal health management that maintains private, updatable memory across repeated encounters, substantially improving answer accuracy and privacy while reducing prompt context exposure.

• Separates shared safety rules/medical knowledge from private longitudinal memory (profile facts, reusable procedures, episodic traces); after each episode an induction step decides what to update, revise, keep episodic, or exclude.
• On 900 longitudinal support probes, accuracy rose from 0.2% (current-query prompting) to 45.7%, with 71.7% lower prompt-side context exposure than full-history prompting.
• In 100 privacy probes it gave higher privacy-aware answer quality and fewer unsafe disclosures than both baselines; across nine 200-case biomedical tasks, mean absolute gain was 27.0 points, with seven gains significant after FDR correction.
• Results are offline/synthetic-benchmark based; the authors note clinical effectiveness needs prospective evaluation.