ECHO: Locally-Deployable AI Health Assistant with Temporal Memory and Safety Guardrails
So, there's this new study on arXiv (2608.06110) that introduces ECHO, which stands for Enhanced Care & Health Observer. It’s a health assistant designed for managing chronic conditions, and it can work locally. The system consists of three connected software modules all working under a single framework. The main feature is a chatbot that uses a ReAct loop powered by LangGraph. It includes 17 clinical tools and keeps track of information through a temporal knowledge graph. It achieved a 94.9% success rate in executing tasks across 59 scenarios with GPT-5 Mini. The safety system processes queries in two steps: a rule-based system for urgent issues and a GNN that accurately handles tricky cases, scoring 88.8% accuracy. The study highlights the system's design and its potential for secure, on-device health support.
Key facts
- ECHO is a locally-deployable conversational health assistant for chronic care management.
- It integrates three software modules under shared supervision.
- Core module: agentic chatbot using ReAct loop via LangGraph, with 17 clinical tools and temporal knowledge graph.
- Achieves 94.9% tool-execution pass rate on 59-scenario benchmark with GPT-5 Mini.
- Two-stage hybrid safety layer: rule-based layer handles crisis signals and jailbreaks in <1ms.
- Signed GNN with APPNP-style propagation classifies boundary cases with 88.8% accuracy and 90.6% unsafe recall.
- Evaluated on 2,537-query annotated Turkish health dataset.
- Outperforms zero-shot LLM baselines.
- Paper available on arXiv with ID 2608.06110.
Entities
Institutions
- arXiv