Search papers, labs, and topics across Lattice.
This paper introduces HealthClaw, an open-source agent architecture designed for longitudinal personal health management that evolves its support based on changing user routines and preferences. By separating safety rules and medical knowledge from private longitudinal memory, HealthClaw significantly improves answer accuracy from 0.2% to 45.7% across 900 longitudinal support probes, while also reducing prompt-side context exposure by 71.7%. Evaluations show that HealthClaw not only enhances answer quality but also maintains higher privacy standards compared to existing systems, indicating its potential for safer and more effective health management solutions.
HealthClaw boosts answer accuracy for personal health management by over 45% while enhancing privacy protection in AI interactions.
Personal health management unfolds over repeated encounters, yet most health AI systems treat each request in isolation. We developed HealthClaw, an open-source agent architecture that updates support as a person's routines, preferences, measurements and risks change. It separates shared safety rules and medical knowledge from private longitudinal memory containing profile facts, reusable procedures and episodic traces. After each episode, induction determines what should update the profile, revise a procedure, remain episodic or be excluded. We evaluated HealthClaw with a synthetic year-long benchmark and nine 200-case biomedical tasks. Across 900 longitudinal support probes, answer accuracy increased from 0.2% with current-query prompting to 45.7% with HealthClaw, while prompt-side context exposure was 71.7% lower than with full-history prompting. In 100 privacy probes, HealthClaw produced higher privacy-aware answer quality and fewer unsafe disclosures than both baselines. Across the biomedical tasks, the mean absolute gain in the task-specific primary metric was 27.0 percentage points, and seven gains remained significant after false-discovery-rate correction. These offline benchmarks support governed, self-evolving memory for longitudinal personal health agents, although clinical effectiveness requires prospective evaluation. HealthClaw is publicly available at https://github.com/HC-Guo/HealthClaw.
