API reference
Conversations and history
How conversation_id works, how much history the model sees, and where conversations from every channel show up.
Reference1 min readUpdated 23 October 2026
SentientOne keeps the history of every conversation for you. Send a message without a conversation_id and a new conversation starts; the meta and done events give you its id. Send that id with the next message and the agent sees the earlier turns. You never resend the history yourself.
{ "message": "And when will it arrive?", "conversation_id": "9b1c…"}How much history the model sees
| Value | |
|---|---|
| Messages sent to the model | The last 12 messages, up to 12,000 characters in total |
| Messages kept per conversation for context | Up to 20, each cut to 4,000 characters in the fast cache |
| Repeated messages | A user message that repeats the previous assistant reply word for word is dropped from the context |
Long conversations keep working: older turns simply fall out of the model's view. If something from early on matters, have your app restate it, or put it in the agent's knowledge.
Rules
- A conversation belongs to one agent. Sending another agent's conversation id fails with
CONVERSATION_NOT_FOUND, sent as anerrorevent on the API and as404on the widget. - An id that doesn't exist fails the same way. Start a new conversation by leaving the field out.
- Each request is still one message and one credit, however long the conversation is.
- Continuing a conversation always skips the answer cache.
Where to read them
Every conversation from every channel (API, widget, Playground) is on the agent's Conversations tab. Each reply also has a trace with what was retrieved and every tool call: see Tracing.
Does the widget keep the conversation if the visitor reloads the page?
No. The widget holds the conversation id in memory, so a reload starts a new conversation.
Can I delete a conversation?
Conversations are deleted with their agent, or with your account. See Delete your account.
Still stuck?
Email us with what you tried and any request id. A person replies.