It's stored associated with only a randomised ID as metadata, and it loses that association after six months. It then survives untethered for the remaining eighteen.
I'd therefore posit that it's being used for Siri training, not for the sake of "acquiring personal behavioural data".
That is to answer your question literally. It doesn't make me a lot happier to know that my requests would be stored for two years even if they didn't have details attached that more explicitly identify me. That's partly why I don't use Siri, though my main justification is the same as for the other voice assistants: I'm unclear of any situation in which they'd be useful for me.