Prodigy sends the answered questions back in batches – usually the same as the stream batch size, which defaults to 10. To allow being able to quickly undo answers (without having to reconcile existing annotations), the latest 10 answers are also kept in the app. As soon as there’s a full batch of answers minus the undo history, the app will send those answers back to the server. This also makes it easy to update a model in the loop with batches of answers.
The saving happens automatically in the background – but before exiting, you usually always want to hit “save”, to explicitly tell Prodigy you’re done and send everything that’s there back to the server at once.