Skip to main content
Memory lets your assistants remember durable facts about an end user across conversations — a stated preference, an account detail, a past issue — so every interaction picks up where the last one left off. Feather extracts these facts as conversations close and recalls the relevant ones at the start of future turns.
Memory is backed by Feather’s in-house facts store (PostgreSQL). It is the sole memory backend — there are no third-party memory vendors.

How memory is scoped

Facts are scoped to a (organization, end_user) pair — one memory thread per person, persisting across individual conversations. When a conversation contributes a fact, Feather records which session it came from (source_session_id) for attribution, but the fact itself belongs to the end user, not the session. Memory is governed by an organization-level toggle. Turn it on or off with GET/PUT /v1/identity/org/memory ({ "memory_enabled": true }). When it’s off, no new facts are ingested and recall is suppressed — but existing facts remain readable and deletable. You can also toggle memory per assistant revision with the memory_enabled field.

Reading memory

A fact looks like:
The session view also reports a state of pending, synced, skipped, or suppressed, so you can tell whether a conversation’s facts have been ingested yet.

Deleting memory

Memory deletion is immediate and scoped to the org and user.
Deleting a user’s memory removes the derived facts. It is not a full right-to-be-forgotten erasure of the end user’s identity and transcripts — for that, use the identity and privacy tooling.

Next steps

Conversations

How turns, closing, and post-session processing produce memory.

Privacy & data protection

Control PII scrubbing and data retention across memory, transcripts, and audio.