PIA: A Personal Intelligence Agent Turning Health Conversations into Records and Records into Understanding

Authors: Jeonghun Yoon, Dongchan Kim, Hongyeon Yu, Young-Bum Kim, Jaegul Choo
Preprint: arXiv:2609.31255 [cs.AI, cs.CL]
Primary Subjects: Artificial Intelligence (cs.AI), Computation and Language (cs.CL)
Direct Links: arXiv Abstract | PDF Fulltext

Abstract

General-purpose agent memory summarizes conversations: it extracts salient snippets, embeds them, and retrieves the top-$k$ into the prompt. A health agent cannot run on summaries: a dose becomes a sentence, “since last week” is resolved at the model’s discretion, and a three-month glucose trend cannot be answered by text similarity alone.

We present PIA, a personal intelligence agent deployed alongside a consumer health agent. PIA receives the agent’s natural-language requests, decides for itself whether and how to write or read, and turns conversations into typed clinical records and records into a synthesized understanding of the user. Its memory harness consists of four controls—extraction, memory, retrieval, and understanding—each a domain-agnostic mechanism paired with a pluggable health module: schema, medical alias dictionary, knowledge graph, and temporal rules.

We show how the same query receives a different answer as the memory injected into the response context deepens from one-dimensional recall, to a two-dimensional health snapshot, to a three-dimensional trajectory with causality, and report lessons from operation: self-reported health data are missing not at random, question phrasing governs the quality of synthesized understanding, and nearly a third of candidate causal links are structural noise that rules alone remove.

Research Problem & Motivation

Standard Large Language Model (LLM) agents rely on generic episodic memory schemes:

• Summarization Discarding Numerical Precision: Reducing a clinician consultation or patient dialogue to text summaries flattens vital therapeutic details (e.g. converting “Metformin 500mg twice daily with meals” to “user takes diabetes medication”).
• Loss of Temporal Grounding: Relative references (“fever recurred three days after starting antibiotics”) lose deterministic timestamp anchoring once summarized.
• Vector Similarity Blindness: Cosine distance in embedding space cannot evaluate time-series dynamics, volatility, dosage titration, or multi-month metabolic trajectories (e.g., fasting blood glucose trends).

System Architecture: The Four-Control Memory Harness

PIA solves conversational memory degradation by establishing a formal Memory Harness governed by four specialized operational controls:

[Raw User Dialogue / Consultations]
                │
                ▼
  [1. Extraction Control: Schema Module]
  ──> Converts free-form text into strongly typed, clinically validated entries
                │
                ▼
  [2. Memory Control: Medical Alias Dictionary]
  ──> Resolves patient phrasing & trade names to canonical ontology IDs (RxNorm, SNOMED)
                │
                ▼
  [3. Retrieval Control: Knowledge Graph Module]
  ──> Constructs episodic graph links between medications, symptoms, and lifestyle
                │
                ▼
  [4. Understanding Control: Temporal Rules Engine]
  ──> Performs deterministic trajectory computation & causal relationship synthesis
                │
                ▼
[Synthesized Longitudinal Clinical Understanding]

1. Extraction (Schema Module): Extracts typed attributes (entity name, quantitative dose, frequency, route, date range, anatomical site).
2. Memory (Alias Normalization): Resolves ambiguity across patient brand names, generic formulations, and abbreviations.
3. Retrieval (Knowledge Graph): Connects multi-session episodic observations into a queryable graph structure, enabling multi-hop associative queries across conditions, medications, and lifestyle changes.
4. Understanding (Temporal Rules Engine): Formulates deterministic time-series queries and chronological trend models to derive longitudinal insights without hallucination.

Key Empirical Findings & Operational Lessons

• Memory Depth Progression: As agent context deepens across dimensions:
  1. 1D Recall: Pure lexical/embedding retrieval returns isolated factual snippets.
  2. 2D Snapshot: Tabular cross-sectional status provides current regimen state.
  3. 3D Causal Trajectory: Longitudinal graph modeling uncovers causal links (e.g., adverse drug effects vs. symptom onset).
• Non-Random Missingness: In real-world self-reporting, data entries are non-randomly missing. PIA’s schema harness prevents structural noise from contaminating the long-term context window.
• Pruning Structural Noise: Unconstrained LLM graph induction hallucinates spurious causal relationships; applying deterministic temporal validation rules prunes ~30% of false causal candidates.

Citation

@article{yoon2026pia,
  title={PIA: A Personal Intelligence Agent Turning Health Conversations into Records and Records into Understanding},
  author={Yoon, Jeonghun and Kim, Dongchan and Yu, Hongyeon and Kim, Young-Bum and Choo, Jaegul},
  journal={arXiv preprint arXiv:2609.31255},
  year={2026},
  url={https://arxiv.org/abs/2609.31255}
}