Memory
How the Clinia Context Engine captures conversational statements, keeps them separate from the structured clinical record, and links them back to facts through typed relationship edges.
Memory
A memory is a piece of unstructured information derived from a conversation with a patient, a caregiver or a provider. It represents something said during a chat session that is worth keeping, but that is deliberately not promoted into the canonical structured clinical record because a conversation is lower-trust than ingested clinical sources (FHIR, CDA).
Two kinds of memory
Memories split into two kinds based on whether they relate to the structured clinical record.
type | What it captures | Links to facts |
|---|---|---|
symptom | Self-reported clinical state ("headaches every morning for two weeks") | Yes — condition facts |
adherence | Medication-taking behavior ("I stopped taking X", "I skip doses") | Yes — medication facts |
concern | A worry, fear, or open question raised by the patient | No |
preference | Communication, scheduling, care, or identity preferences | No |
symptom and adherence are clinically-anchored: they relate back to existing facts and produce relationships when memories are generated. concern and preference are standalone: they are recorded as-is with no relationships to the structured patient record graph.
Append-only
Memories are never deduplicated, merged, or updated. Every new conversational statement is recorded as-is. This is the opposite of facts, which are deduplicated and merged across all ingested sources.
Enriching the record without changing it
A memory never mutates the fact it relates to, nor is it ever promoted to a fact. If a patient says something that contradicts their structured record, that disagreement is expressed as a conflicts relationship from the memory to the fact. The fact's value, status, and confidence are left untouched. symptom and adherence memories can also produce corroborates or enriches relationships when the statement agrees with or adds context to the record.
The linkage between a memory and a fact is always expressed as a typed relationship, never as a direct reference on the memory itself. See Relationship Inference for how relationships are typed and stored.
When memories are created
Extraction is triggered on demand by calling the generate-memories endpoint on a session, not by any automatic session-end event. When triggered, the session's accumulated transcript is analyzed to identify statements matching the four memory types. For clinically-anchored memories, relationships to matching facts are inferred in the same pass. See Sessions for the full request flow, and Manage a Session for a worked example.
Where memories show up
Every memory is readable through the VFS at /memory/{type}.md, one file per type, most recent first. See VFS Paths for the full path reference.
Clinically-anchored memories also surface as a "Patient-Reported" section on the fact they're related to: symptom memories on the matching condition's condition story, and adherence memories on the matching medication's file (/medications/{status}/{slug}, see VFS Paths).