How Interactions Become Memory
How the Context Engine learns memories from conversations, keeps them apart from the structured record and links them to the facts they speak to.
The Context Engine forms Learned Memory from agent interactions. Learned memories hold what is learned about a patient during a session. The engine attributes each learned memory to its author: the patient, a caregiver or a clinician.
What a memory is
A memory is a piece of information the engine keeps from a conversation, stored close to how it was said, such as a symptom, a skipped dose, a concern or a preference.
How memories stay apart from the record
Learned memories are append-only. Each one is kept as it was said and is never merged, deduplicated or updated, unlike facts.
A learned memory can be linked to the facts it speaks to. The link shows the memory next to the fact as extra context and never changes the fact, even when the memory contradicts it. The memory types, link types and the rules they follow are listed in From Theory to Practice.
When memories form
Your application sends a conversation's turns to POST /v1/patients/{patientId}/ingestions/sessions, either turn by turn as the conversation happens or as a full transcript after it ends. Each call triggers an ingestion.
Where memories show up
An agent reads a patient's memories in the /memory/ folder of the virtual file system, one file per type with the most recent first, as described in VFS Paths. Linked memories also appear next to the fact they speak to.
Next steps
- Sessions for how a session groups the turns of one conversation.
- Manage a Session to record a conversation turn by turn.
- Ingest Sessions to import transcripts of conversations that already happened.
- How the Engine Thinks About Data for the model behind Learned Memory.
- From Theory to Practice for the objects the API returns.