Skip to main content
Text and voice share the Conversation Gateway. It assigns and normalizes message/session context, preserves finalized transcripts, and sends only bounded provider work. Duplicate event handling is idempotent; incomplete or interrupted audio does not become a confident teaching. T10 proved voice teach → fresh text recall → text correction → voice recall. T11 proved reconnect, idempotent replay, and provider failure preservation. Raw audio is transient; finalized transcripts are the retained representation.