A temporary chat is a full LADLE conversation that is never persisted to the database. No chat row, no message rows, no artifacts, no share-ability, no history. The connection to Anthropic is real — you get real replies from the real model with the real system prompt — but nothing about the conversation survives past the browser tab being closed. Ideal for one-shot lookups you don't want in your history.
Starting a temporary chat
Three ways to start a temporary chat. Each results in the same server-side behavior — chat_id stays null, persistence is skipped, meal accrual still fires.
What is not saved
Every user-owned surface is skipped. The connection to the model is real; the persistence layer is bypassed entirely. Nothing survives the tab close.
What still fires
The meal accrual counter increments. Your subscription still funded whatever fraction of a meal the request cost — a temp chat contributes to the ledger the same way any other chat does. Usage is recorded too, so this cycle's compute allowance reflects the temp turn.
UI treatment
A persistent amber banner sits at the top of the temp-chat surface reading 'TEMPORARY CHAT · NOT SAVED'. The assistant is instructed via the system prompt to answer normally but to skip any 'as we discussed earlier' or 'based on our previous conversation' phrasing — there is no previous conversation to reference.
When temp chats are the right tool
Quick factual lookups you don't want cluttering your history. Testing a prompt formulation without polluting your project. Asking about something private you'd rather not have in your export. One-off questions to check if the model knows something. Not appropriate for: iterative work you might want to come back to; anything that produces an artifact worth keeping; anything you want the assistant to remember. If you're not sure, don't start it as a temp — you can always archive a persistent chat later.
Docs: Memory and instructions