---
title: "Longitudinal memory for healthcare AI agents"
description: "Healthcare AI agents need longitudinal context: what changed across visits, the source note behind every fact, and an insufficient-evidence status when the record is silent."
canonical: https://past.dev/industries/healthcare
last-updated: 2026-08-31
---
# Memory infrastructure for healthcare AI

Source: https://past.dev/industries/healthcare

Healthcare AI agents often work from individual encounter records. Longitudinal questions require facts and changes across visits, with sources for each claim. past.dev stores dated facts across encounters, reports conflicting records and returns `NoKnownSupport` when the supplied history does not establish an answer.

## The memory problems inside healthcare AI

- **Longitudinal context.** Encounter tools produce per-visit artifacts. Questions arrive across visits: what changed since March, when did the symptom first appear, which medication was current at the time.
- **Conflicting records.** The chart says one thing, the patient reports another, a second system disagrees. In clinical work a conflict is information that must surface, never be averaged away.
- **Evidence per claim.** A clinical statement without its source note is not actionable. Every recalled fact needs the document and date behind it.
- **Unsupported questions.** The application needs an explicit insufficient-evidence result when the record does not establish an answer.
- **Backfilled histories.** Importing years of records must preserve original event dates, or every timeline question silently breaks.

## How clinical workflows use recall status

past.dev recall returns ranked, dated evidence with a status value. A healthcare application can use `Supported` to answer with source-note citations, `Conflicted` to present both records and their dates, and `NoKnownSupport` to report that the record does not establish an answer.

Time is equally structural: facts carry [validity windows](/glossary/validity-window), so an agent can answer what was current at the time of a past visit, and change questions derive from [supersession](/glossary/fact-supersession) records rather than from re-reading every note.

## How the integration works

Agents send visit summaries, messages, and intake notes through `POST /api/v1/ingest` with each record's original timestamp. Entity resolution links the same patient, clinician, and medication across sources. `POST /api/v1/recall` returns dated evidence and a status value. The [quickstart](/docs/memory-api/quickstart) documents the four API calls. The [Claude Agent SDK](/integrations/claude-agent-sdk) and [LangGraph](/integrations/langgraph) guides show framework integrations.

## Security posture and regulatory reality

The [security page](/security) documents SOC 2 Type II, ISO frameworks, encryption, deletion and retention behavior. HIPAA obligations remain with covered entities and their vendors. Infrastructure can provide controls, audit logs, deletion and agreements, but cannot make an application compliant by itself. HIPAA has no product certification. For deployment and agreement questions, [talk to us](/contact).

Recall quality, including conflict handling and abstention, is measured in public with the method on the [benchmarks page](/benchmarks).

## Frequently asked questions

### Does past.dev store PHI?

past.dev stores the text you send it. Whether that includes PHI is an architectural choice you control: some teams send full clinical text, others send de-identified or reference-linked records. Review the security page for controls and talk to us about the agreements your deployment needs.

### How does this relate to an EHR or FHIR store?

It does not replace either. The EHR stays the system of record. past.dev is the queryable memory over the unstructured stream around it, messages, notes, summaries, with time and provenance, so agents can answer longitudinal questions the chart snapshot cannot.

### What happens when two records disagree?

Both are retained and recall returns Conflicted with the evidence and dates on each side. Silent resolution is the failure mode clinical workflows cannot afford, so it is not the default.

## Related

- [Fact supersession](https://past.dev/glossary/fact-supersession)
- [Validity window](https://past.dev/glossary/validity-window)
- [Memory for accounting AI](https://past.dev/industries/accounting)
- [What is agent memory?](https://past.dev/what-is-agent-memory)
- [Quickstart](https://past.dev/docs/memory-api/quickstart)
- [Security and certifications](https://past.dev/security)