Skip to content

A Self-Evolving Agent for Longitudinal Personal Health Management

🕒 Published (v1): 2026-07-15 15:22 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

HealthClaw is an open-source LLM agent architecture for longitudinal personal health management that uses a five-layer memory system with post-episode induction to selectively update user-specific state rather than appending raw history. It achieves a 45.7% answer accuracy on year-long support queries (vs. 0.2% for current-query-only prompting) while reducing prompt context by 71.7% versus full-history prompting.

Problem

Most health AI systems treat each user interaction in isolation; they lack mechanisms for selectively updating personal state (allergies, routines, medication history, preferences) across months of repeated encounters. Naively supplying full history at each turn improves recall but increases context cost, introduces noise from irrelevant prior exchanges, and exposes sensitive health data unnecessarily.

Method

HealthClaw implements a closed perception–reasoning–action–induction loop over a five-layer memory hierarchy: - L0: shared behavioural/safety rules; L1: shared domain knowledge and tool-routing index (both static, shared across users). - L2: privacy-critical personal profile (chronic conditions, allergies, stable preferences) stored locally; L3: reusable task SOPs (meal planning, risk routing, privacy disclosure); L4: episodic traces.

After each episode, a post-episode induction step classifies each item from the interaction into one of four fates: (1) update L2 profile facts, (2) revise an L3 SOP, (3) retain as an L4 episodic trace, or (4) exclude from future use. Retrieval is selective—only task-relevant memory layers are surfaced, preventing wholesale exposure of the longitudinal record. Privacy-sensitive facts can be withheld from writeback entirely.

Key Contributions

  • Five-layer memory architecture separating shared governance (L0–L1) from user-specific longitudinal state (L2–L4), with per-layer update and retrieval policies.
  • Post-episode induction mechanism that assigns each interaction item a distinct persistence fate instead of appending raw dialogue to future prompts.
  • A synthetic 365-day longitudinal benchmark with 900 support queries and 100 privacy probes drawn from simulated daily trajectories.
  • Demonstrated privacy-governed memory writeback: unauthorized disclosure rate of 5% vs. 18% (current-only) and 15% (full-history).
  • Evaluation across nine heterogeneous biomedical tasks (imaging, EHR, physiological signals, genomics, proteomics, multi-omics), showing a mean absolute gain of 27.0 pp in task-specific primary metric.
  • Open-source release of the full HealthClaw architecture.

Results

Longitudinal support (n = 900 queries): - Rubric-defined answer accuracy: HealthClaw 45.7% vs. current-only 0.2% vs. full-history 61.2% - Automated rubric score: 0.568 vs. 0.182 vs. 0.752 - Reference-fact coverage: 0.524 vs. 0.027 vs. 0.759 - Mean prompt characters: 18,274 (HealthClaw) vs. 64,493 (full-history) — 71.7% reduction

Privacy probes (n = 100): - Answer accuracy: HealthClaw 64.0% vs. current-only 44.0% vs. full-history 40.0% - Constraint violations: 24% vs. 41% vs. 53% - Unauthorized disclosure: 5% vs. 18% vs. 15% - Privacy-preserving alternatives offered: 51% vs. 27% vs. 18%

Biomedical tasks (9 tasks × 200 cases): - Mean absolute gain over base condition: 27.0 pp - 7 of 9 tasks significant after Benjamini–Hochberg FDR correction - Largest gains: NoduleMNIST3D (+61.5 pp), GeneTuring (+52.5 pp), PAD-UFES-20 (+41.5 pp) - Non-significant gains: PhysioNet ICU SOFA (+5.0 pp), diabetes readmission (+4.5 pp accuracy)

Limitations

  • All longitudinal trajectories are synthetic; evaluation uses a Qwen-3.7-based automated rubric with no human ratings.
  • Biomedical task comparisons use different model backbones and evaluation pipelines per task, not controlled same-model ablations; some task-specific tools used resources close to benchmark sources.
  • Full-history prompting still outperforms HealthClaw on non-privacy longitudinal recall (61.2% vs. 45.7%), indicating the selective memory mechanism trades off some recall for context efficiency and privacy.
  • One failure case (BEHSOF fatty-liver screening): memory-retrieved evidence reinforced an existing model prior (NAFLD bias), narrowing class predictions—suggesting risks of confirmation bias in iterative memory accumulation.
  • No prospective clinical evaluation; privacy guarantees are constrained to the agent loop and do not encompass deployment-level access control or security.
  • Induction step correctness is not formally verified; noisy or sparse feedback may degrade writeback quality over long horizons.

Relevance to Agentic AI / LLM Agents

HealthClaw directly addresses a core open problem in long-horizon agentic systems: how to maintain and selectively update a user's persistent state across many episodes without either discarding context (stateless) or accumulating it indiscriminately (full-history). The five-layer memory with post-episode induction is a concrete instantiation of memory governance—an architectural pattern applicable beyond health to any agent that must serve the same user over time. The work also highlights the tension between recall and privacy/context-efficiency that any persistent LLM agent faces, and provides a structured empirical framework (longitudinal benchmark + privacy probes) for evaluating it. It connects self-improving agent research (Reflexion, ExpeL, Voyager) with domain-specific deployment constraints in a setting where memory errors carry real-world risk.