Smart Memory
Turn onmemory to recall past conversations with the same user.
- Each turn: the server searches the user’s memory and adds the relevant context to the prompt. The model uses it without reading out the memory system.
- At close: the session transcript is saved to the user’s memory, so the next session can use it.
- Permission: the API key needs the memory permission. Otherwise the session continues without memory.
- Scope: memory is per user. Memories show in the Memory pages and the Memory API.
Each memory or knowledge lookup that adds context to a reply costs $0.0003. A turn with nothing relevant found is not charged.
Knowledge collections
Point a session at Memory collections so answers come from your documents.- Up to 5 collection IDs per session.
- Create and upload collections in Memory → Collections, or with the Memory API.
Intent filler and hints
intent lets a small decision model choose a fitting spoken acknowledgement while the main reply is written. This hides latency on questions that take a moment.
- No filler is spoken for greetings or goodbyes.
Multi-language transcription
Pass an array of language codes to transcribe a mixed-language conversation:- A single string such as
"en"is also accepted. - When
languageis omitted, the default is English and Hindi. - The reply uses the user’s language, in its native script.