A Self-Evolving Agent for Longitudinal Personal Health Management

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Haoran Li, Jiebi Deng, Tong Jin, Jinghong Han, Yuxin Wang, Zexin Wang, Qingyi Si, Weikang Gong, Xiahai Zhuang, Jia You, Wei Cheng, Jianfeng Feng, Hongcheng Guo
• Venue: arXiv: cs.AI
• Published: 2026-07-15
• Collected: 2026-07-16
• URL: http://arxiv.org/abs/2607.13940v1
• PDF: https://arxiv.org/pdf/2607.13940v1
• DOI: UNKNOWN

Plain-Language Brief

The authors developed an open-source agent architecture called HealthClaw for personal health management. It learns and adapts to changes in a person's routines, preferences, measurements, and risks over time. The system was evaluated through synthetic benchmarks and biomedical tasks, showing improved answer accuracy and privacy compared to current systems.

What the Document Says

• HealthClaw is an open-source agent architecture for personal health management.
• 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.
• The system was evaluated with a synthetic year-long benchmark and nine biomedical tasks involving 900 longitudinal support probes and 100 privacy probes.
• HealthClaw showed improved answer accuracy compared to current systems, with lower prompt-side context exposure.
• In privacy probes, HealthClaw produced higher privacy-aware answer quality and fewer unsafe disclosures than baselines.
• Across biomedical tasks, HealthClaw had a mean absolute gain of 27.0 percentage points in the task-specific primary metric, with seven gains remaining significant after false-discovery-rate correction.
• The offline benchmarks suggest that governed, self-evolving memory is beneficial for longitudinal personal health agents, but clinical effectiveness requires prospective evaluation.
• HealthClaw is publicly available at https://github.com/HC-Guo/HealthClaw

Key Concepts Mentioned

• HealthClaw
• Personal health management
• Longitudinal memory
• Induction
• Synthetic benchmarks
• Biomedical tasks
• Answer accuracy
• Privacy-aware answer quality
• Unsafe disclosures
• Task-specific primary metric
• False-discovery-rate correction

Claims Made by the Document

• HealthClaw improves answer accuracy compared to current systems.
• HealthClaw has lower prompt-side context exposure than some systems.
• HealthClaw produces higher privacy-aware answer quality and fewer unsafe disclosures than baselines.
• HealthClaw had a mean absolute gain of 27.0 percentage points in the task-specific primary metric compared to other systems, with seven gains remaining significant after false-discovery-rate correction.

Evidence Used

• Increase from 0.2% with current-query prompting to 45.7% with HealthClaw
• 71.7% lower than with full-history prompting
• HealthClaw produced higher privacy-aware answer quality and fewer unsafe disclosures than both baselines
• Mean absolute gain in the task-specific primary metric was 27.0 percentage points, and seven gains remained significant after false-discovery-rate correction

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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.