Memory

Memory is what lets an agent remember a person across conversations: their preferences, past issues, and the facts that matter. It's distinct from knowledge: knowledge is your content; memory is what the agent learns about the people it talks to.

How memories form

As conversations happen, the agent records durable facts as memories. Each memory has content, a confidence score (0–1), and an optional quality score (1–5), and is linked to the person (customer_id), and, where relevant, the source conversation it came from.

Scope, category & tier

Memories are organized so the right ones surface at the right time:

• Scope - temporary or permanent.
• Category - semantic (facts) or episodic (events), with an optional event_timestamp and extracted entities.
• Tier - working, short_term, long_term, or archival, reflecting how central a memory is.

Memories also track usage (last_accessed_at, access_count) and can carry an expires_at.

Review, verify & correct

You stay in control of what an agent believes. Review memories in the dashboard or over the Memories API; mark a memory verified, adjust its confidence, edit its content, or delete it. To audit what's been learned, the Memory page lets you search the text and narrow it down by agent, type, or verification status. The API can also filter by category.