Typed memory with provenance, for regulated environments.
Healthcare is a vertical Flowlines is designed for, not one we are deployed in today. This page explains how the architecture maps to clinical agents: typed fields, full provenance, and signals for missing critical context.
The problem
A patient's conditions, medications, and allergies need to be tracked across sessions. Free-text conversation history buried in a context window is not the same as a typed, scoped field the agent can reason about, or that a reviewer can audit.
Every piece of clinical context that influences an agent response needs a trail back to the interaction that produced it. Provenance is a prerequisite for review in any regulated setting, and it's the first thing that disappears when memory is a vector blob.
The architecture
Flowlines writes structured fields (conditions, medications, allergies, contraindications) each linked to the source turn. Signals fire when a required field is empty, when output contradicts recorded state, or when a recommendation conflicts with a known constraint.
What it catches
Contraindication risk: a recommendation that conflicts with a recorded allergy or condition, flagged before delivery. Context contradiction: the agent stating something that disagrees with previously recorded patient information. Missing critical field: a required field with no value, which triggers structured intake instead of a guess.
How deployment works
Self-hostable on your cloud. PII redaction by default. Human-in-the-loop is the default; the system surfaces, your team decides. Nothing writes to memory without a traceable source interaction.
Structured memory fields
The typed fields Flowlines reads and writes for this domain. Each field is scoped, versioned, and traceable back to the interaction that produced it.
conditionsmedicationsallergiescontraindicationsconsent_scope