Technical Summary: HiMe - Real-Time Self-Hosted Personal Agent Platform for Health Insights

1. Problem Statement

Traditional wearable health analytics (e.g., smartwatches) are constrained by rigid, fixed statistical frameworks that fail to capture signal dynamics or user preferences. Furthermore, the longitudinal raw data streams generated by these devices are too voluminous to fit directly into the context windows of Large Language Models (LLMs) for intelligent analysis. While LLM agents offer a promising alternative for reasoning over health data, existing solutions fall short in three key areas:

• Clinical systems focus on encounter-based records for clinicians, not continuous personal data for patients.
• Personal coaching systems rely on guided dialogue or journaling rather than continuous physiological streams.
• General-purpose self-hosted assistants are designed for chat and scheduled tasks, lacking the capability to ingest high-rate physiological data and model users over long horizons.

There is currently no open-source, locally deployable platform capable of processing personal health data in real-time while preserving privacy and generating long-term, personalized health insights.

2. Methodology and Architecture

The authors present HiMe (Health Intelligence Management Engine), a self-hosted, privacy-first agent platform designed to run on user-owned hardware. The system architecture is defined by three core design principles and a four-layer framework.

Core Design Principles

1. Database as a First-Class Component: The database is not merely a storage backend but a central, checkable terminal state. This allows the agent to jointly analyze wearable signals, user preferences, and memory. It also enables automated evaluation by comparing the agent's actions against a "gold state" in the database.
2. Joint Optimization of Effectiveness and Efficiency: The system seeks a Pareto-optimal balance between quality, cost, and latency. This involves using lightweight statistical filters to monitor data streams at full resolution, waking expensive LLM inference only when specific triggers fire.
3. Real-Time Processing with Long-Term Modeling: The system captures both momentary anomalies and months-long trends, supporting timely intervention and longitudinal understanding.

System Architecture (Four Layers)

• Data Layer: A unified per-user database stores wearable signals, memory, user profiles, reports, and agent-created tables. This allows heterogeneous data to be analyzed jointly. The agent has specific table permissions to ensure safety.
• Tool Layer: The agent interacts primarily through read/write operations on the unified database and a code execution tool (maintaining Jupyter-style state). A "fact verifier" classifies outbound messages, checks reported numbers against query evidence, and stores hashed evidence trails to prevent hallucinations.
• Agent Layer: A main Interaction Agent routes user messages to specialized Sub-Agents:
  • Analysis Agent: Answers natural language questions over data.
  • Management Agent: Handles memory updates and proactive tasks.
  • Plan Agent: Generates personalized health plans.
    Prompts are modular (system instructions, role, soul, etc.) to maximize KV-cache reuse.
• Interaction Layer: Users interact via built-in chat or integrated messaging platforms (e.g., WeChat, Telegram). The system supports proactive notifications and features a "Digital Twin" (a pixel-art cat) that reflects the user's health status in real-time.

Workflow

Wearable devices stream signals (heart rate, SpO2, sleep, etc.) to a smartphone, which buffers and relays them to the local agent server. A data adapter normalizes formats and deduplicates samples. The agent server runs a loop interleaving user messages, real-time monitor triggers, and scheduled tasks. A cheap statistical trigger monitors the stream continuously; when a condition is met, it wakes the expensive analysis agent to generate insights, which are then pushed back to the user's devices.

3. Key Contributions

• HiMe Platform: The first open-source, locally deployable platform that ingests real-time wearable data and generates long-term personal health insights while keeping data on user hardware.
• Measurable Agent Evaluation: By treating the database as a first-class component with a checkable terminal state, the authors enable automated, oracle-based evaluation of agent capabilities without relying on subjective LLM judges.
• Hybrid Monitoring Strategy: The integration of continuous, low-cost statistical monitoring with on-demand, high-cost LLM analysis to balance privacy, latency, and cost.
• Comprehensive Benchmarking: The system was evaluated across five public wearable datasets (LifeSnaps, PMData, MMASH, CovIdenify, GLOBEM) and 22 different model backbones (ranging from 1.5B to 35B parameters, including local and hosted APIs).

4. Results

The evaluation focused on per-role capabilities, composite interaction reliability, and real-time triggering efficiency.

• Local vs. Hosted Models: Strong local models (e.g., Qwen3.6-27b, Qwen3.5-35b-a3b) achieved analysis scores of 0.91, matching or surpassing hosted frontier models like DeepSeek-V4-flash (0.85) and GPT-5-mini (0.76) with zero hallucinations. Only Gemini-3.5-flash (0.94) led significantly.
• Model Size vs. Capability: Capability did not scale monotonically with parameter count. For instance, Qwen3-30b-a3b (30B total, ~3B active) underperformed the smaller dense Qwen3-8b on analysis (0.36 vs. 0.47) and memory tasks.
• Small Model Limitations: Models with 1.5B parameters failed across all capabilities, with high hallucination rates (0.47), indicating current small on-device models are insufficient for complex health agent tasks.
• Composite Reliability: While per-turn scores were high (e.g., 0.96 for Qwen3.5-35b-a3b), strict multi-turn reliability (S5 score) remained low (max 0.44), indicating that chaining competent roles across a session is still an open challenge.
• Trigger Efficiency: The "trigger" mechanism (statistical gating) significantly reduced token usage (up to 121x savings for smaller models) while maintaining or improving recall for anomaly detection compared to fixed polling schedules.
• Data-to-Report Narration: Generating natural language summaries from time-series data proved difficult; the best backbone achieved only 46% on a stress/recovery rubric, suggesting that narrating subjective states is a distinct competence from retrieving numbers.
• User Study: A two-month field study with nine participants showed high ratings for Usability (4.67/5) and Proactive Experience (4.56/5). Participants valued the continuous, evidence-backed monitoring, though personalized plan adaptation was rated lower (4.00/5).

5. Significance and Claims

The paper claims that HiMe demonstrates the practical feasibility of privacy-preserving personal health agents running on consumer hardware. By treating the database as a first-class component, the system makes agent behavior measurable and auditable. The authors assert that strong local models are now viable alternatives to hosted APIs for health insights, though challenges remain in multi-turn reliability and the ability to narrate complex subjective states from data.

The platform is positioned not just as a consumer tool but as a research environment for evaluating and training health agents, and for applying large models to real-world wearable data. The authors conclude that while the system is a demonstration rather than a final benchmark, its architecture provides a foundation for developing trustworthy, personalized health agents that help users become "healthier selves over time."