Again, it's not even retained by the device. As it's transiently listening, it's being continually overwritten. When the user uses a wake-up word, there is some machine learning that happens on the device as it adapts the audio model to the speaker to reduce the number of false positives or false negatives for that wake-up word. Then if the user is using Siri, at the point Siri is woken up and being communicated with, that is the initiation of transmission of data to Apple.
On May 29th, 2019. See this statement in context.