cs.CLSep 25, 2026
SavePIA: A Personal Intelligence Agent Turning Health Conversations into Records and Records into Understanding
Organizations: KAIST, Republic of Korea · NAVER Corp., Seongnam, Republic of Korea · NAVER Corp., Bellevue, WA, USA
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. 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 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.
Figures & tables
Figure 1: PIA between a health agent and its user’s history. Each turn is kept twice—as typed records in the system of record and as facts in the associative memory (Hindsight; § 2.1 )—and read through a gate and router that choose structured retrieval, associative recall, or both. HU 2 runs in the background and its output enters every turn through the context pack.
Figure 2: Where HU 2 sits. The system of record decides when to regenerate and supplies provenance; the associative memory supplies the material; the four steps run off the request path; the Health Agent consumes the result through the context pack and never calls HU 2 online.
| Rung | What the memory knows | Status |
|---|---|---|
| 1 What | State summary | production |
| 2 Flow | Long-term user state | prod. / pilot |
| 3 Why | Causal relations | pilot |
| 4 Correction | Recommendation, follow-through | design |
| 5 What-if | Counterfactuals | design |
Table 1: The ladder of understanding and where the deployed system stands.
Figure 3: Depth of answer for one query on a synthetic test user. 1D injects only the memory the question points to: the answer checks the named symptom and drug. 2D adds the health snapshot—static profile and current medication—and ends with a generic caution. 3D adds time: homogeneous transitions link the same coordinate across , , into a trend (dashed); heterogeneous causal transitions link a past cause at another coordinate to the present outcome (solid); inactive nodes carry no edges. Which nodes are active depends on the query. Answer sentences are illustrative: the memory supplies the evidence, the agent writes the sentence.
| Component | Metric | Value |
| Gate, RETAIN | Precision | 100.0 |
| Recall | 89.1 | |
| Accuracy | 94.9 | |
| Gate, READ | Recall | 91.0 |
| Retrieval | Tool selection | 91.0 |
| MRR | 0.785 |
Table 2: How the gate, the read router, and the storage path hold up on the synthetic users. The gate figures come from the 192 router scenarios averaged over four runs, the retrieval figures from the 24 retrieval scenarios over three runs (MRR, recall, and precision are computed on queries that returned at least one result), and the judge figure from the 54 storage scenarios over three runs. All values are percentages except the three ranking metrics.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Control | Domain-agnostic mechanism | Health plug-in (what makes it a health memory) |
|---|---|---|
| Extraction | Schema-constrained extraction; temporal expressions copied verbatim, resolved deterministically; metric trajectories and event containers | 23 HUP types; intake vs. regimen, compound blood pressure, pregnancy-as-status rules; sleep-window and supply end-date arithmetic; checkup results, prescription intakes |
| Memory | Exact-alias admission gate with candidate queue; provenance and confidence per record; knowledge-graph annotation over facts | Versioned medical alias dictionary (319 terms, 4,494 aliases); reference ranges (sex-specific for 11 tests) abnormality; concept tags (condition, panel); drug indications |
| Retrieval | Read gate and router; structured retrieval by type/field; associative recall (Hindsight) + rehydration; time-ordered conflicts; hard boundary on named entities | Concept-expanded queries (“diabetes numbers” fasting glucose, HbA1c; “liver numbers” condition panel); named drugs and tests are exact |
| Understanding | Question contract over consolidated memory; synthesis over recalled memory with provenance; regeneration on record change | Seven health questions; medical-constraint field; risk-factor profile; recommendations confined to one guarded field |
Table 3: The memory harness separates domain-agnostic controls (middle) from health plug-ins (right). Replacing the right-hand column instantiates the harness for another domain; the present paper reports only the health instantiation.
| Layer | Type | What it records (representative fields) |
| Static | sex | biological sex; conditions reference ranges and sex-specific interpretation |
| age | date of birth or age; screening intervals, risk assessment | |
| height | adult height; with weight, BMI | |
| blood type | ABO/Rh | |
| allergies | confirmed drug, food, and environmental allergies; consulted before any drug or diet suggestion | |
| family history | major conditions of first-degree relatives |
Table 4: The Health User Profile: 10 static, 13 episodic, and 5 derived semantic types (translated from the deployed Korean schema).
| Field | Question asked of memory |
|---|---|
| who | In summary, what is this user’s current situation and what are their long-term tendencies? |
| physical_state | How is this user’s physical condition overall, and which changes or findings deserve attention now? |
| behavioral_state | What do this user’s lifestyle and behaviors look like, and which patterns within them are recurring or changing? |
| medical_constraints | What medical constraints must be kept in mind when caring for this user, and what do they mean in daily life? |
| trajectory_goal | What is this user trying to achieve, how has that goal changed over time, and how far have they come? |
| trajectory_preference | What does this user like or avoid, and how does that show in daily life? |
Table 5: The seven questions of the HU 2 contract (translated from the deployed Korean prompts). A shared instruction requests plain prose of 300–500 characters; only the last field may recommend. The shared instruction is given in Appendix E .
| Component | Model |
|---|---|
| Memory gate; read router and planner; compact normalizer and fallback; answer summary | gemma-4-26b-a4b-it |
| Associative memory: fact extraction, causal linking, consolidation | gemma-4-26b-a4b-it |
| Embeddings (retain index, semantic recall) | qwen3-embedding-4b |
| HU 2 memory reasoning (seven questions) | gemma-4-26b-a4b-it |
| Deterministic temporal resolver | no model (rules) |
| Evaluation judge (offline only) | Claude Sonnet 5 |
Table 6: Models per component. The judge is used only in evaluation and never in the serving path.
Figure 4: One health snapshot, two queries. The 2D plane holds the same coordinates for the same synthetic user; each query activates a different subset (filled) and leaves the rest inactive (hollow).
Figure 5: Scenario 2 (synthetic user). Top: the three depths; solid, heterogeneous causal path from the knee-pain report through recent short sleep to the present request; dashed, homogeneous running-distance trend. Bottom: memory available at each depth and its effect.
Figure 6: Scenario 3 (synthetic user). Top: the three depths; dashed, homogeneous lipid trend across two checkups; solid, heterogeneous path from the late-night eating habit to the present lipid values. Bottom: memory available at each depth and its effect.
| Expression (as copied) | Resolution |
|---|---|
| “yesterday” | event date day |
| “three days ago, 80 kg” | event date ; value 80 kg |
| “since last Tuesday” | interval start last Tuesday before ; open end |
| “slept from 1 to 6” | sleep window 300 min; crosses midnight if start end |
| “a ten-day supply, started yesterday” | start ; expected end start 9 days |
| “last month’s checkup” | month month; day empty |
Table 7: Representative deterministic resolver rules (Korean expressions translated).
| Field | Value (active build) |
|---|---|
| standard key | test: + canonical name gongbok hyeoldang ‘fasting blood glucose’ (Hangul); unit mg/dL; category: diabetes tests; LOINC 1558-6; KOSTOM H01897515 |
| aliases | 37 rows, 18 distinct strings, e.g. gongbok hyeoldang ‘fasting blood glucose’; gongboksi hyeoldang ‘blood glucose at fasting’; sikjeon hyeoldang ‘pre-meal blood glucose’; hyeoldang (sikjeon) ‘blood glucose (pre-meal)’; FBS; Fasting blood sugar; Glucose; GLU(FBS); a legacy NLU entity name |
| reference range | one band set, all users, no age split: 100 normal; 100–125 impaired fasting glucose; 126 diabetes range (mg/dL); source: a university hospital’s public test-information page, tier T3 |
| concepts | 9 tags: test panels—diabetes tests (primary, required), metabolic-syndrome tests (primary, required), lipid tests (related); conditions—type 2 diabetes, metabolic syndrome, fatty liver, cardiovascular disease, and two further related conditions |
Table 8: The fasting-glucose entry of the deployed dictionary. Korean aliases are stored in Hangul and shown here in Revised Romanization with an English gloss.