PHIA+: An Evidence-Grounded and Context-Efficient AI Agent Framework for Wearable Data Interpretation

Wearable devices make it easy to collect long-term records of activity, sleep, heart rate, and other behavioral signals. However, translating these data into reliable personal health insights remains difficult. ReAct-style large language model (LLM) agents can compute over wearable records, yet two issues remain: health interpretations may rely on weak evidence, and long sessions accumulate tool traces that increase context pressure. We present PHIA+, an agent framework designed to address these issues. It introduces LLM-guided medical retrieval, which improves evidence reliability by selecting source-aware passages from a medical corpus, and summary-based hierarchical memory, which reduces context pressure by compressing completed reasoning traces while preserving the current one. We evaluate PHIA+ on the PHIA benchmark with objective and open-ended protocols. In the stacked objective question answering (QA) study, hierarchical memory reduces context length by a factor of 11.8 and API token use by 56.8%, while keeping final accuracy not significantly different from full-history reasoning (within 0.22 percentage points). In the open-ended QA study, under five-pass anonymized LLM judging, the RAG-augmented PHIA attains the highest mean medical source reliability score (2.92 out of 5) and the highest mean response quality score (3.84 out of 5) among five retrieval ablations. These results suggest that explicit mechanisms for evidence retrieval and long-session memory can improve source reliability and context efficiency in wearable-health agents, though the reliability of medical retrieval remains bounded by the coverage of the open-access corpus. These findings are useful to developers and researchers building practical wearable-health assistants.