How it works
At the start of a conversation, the AI receives short summaries of a handful of the customer’s most recent conversations, so it knows whether relevant history exists. The summaries alone are never used to answer. To use information from an earlier conversation, the AI has to open it first:- Search scans the customer’s past messages for a keyword, optionally narrowed down by date, channel, or ticket status.
- Read loads the messages of one past conversation in full.
Retrieval never gets in the customer’s way. If past conversations are
temporarily unavailable, the AI continues with what it has in the current
conversation.
What the AI can see
The AI only reaches conversations that belong to the same customer and the same workspace. This works across channels: if a customer wrote an email earlier and now uses the chat, the AI can find the email conversation, as long as both tickets belong to the same customer record. A customer without an identity, such as an anonymous chat visitor, has no history to search. Within a past conversation, the AI sees the messages that were exchanged: the customer’s messages and the replies from the AI or Human Agents. Internal notes, activity events, attachments, and the AI’s own reasoning from the past conversation are never included. Some conversations are left out entirely: deleted tickets, spam, outbound tickets, and test conversations from the Simulator.Safeguards
- When the AI uses information from an earlier conversation, it says so in its reply, so the customer knows where the information comes from.
- Past conversations never replace current, verified data. For sensitive actions, the AI collects and verifies the required information in the current conversation, for example through Trusted-only context variables.
- Access is read-only. The AI never changes past tickets.
- The customer scope is enforced by the platform. The AI cannot request another customer’s or another workspace’s conversations.