I'm not qualified to help you in this area, but I wanted to add that I actually did try to do this once using a much simpler approach (it was kind of a REPL + conditional logic or fact-based rule engine). I called it "Fred".

One thing I thought really interesting and had noticed (and I think you have also) - current digital assistants don't actively/autonomously engage with the user much. Especially not in any way that would make me think they are "friendly". This is definitely something that can go sideways quickly, but on the face of it, I really do love the idea of a benevolent AI that actively tries to interact with me in a safe way (with the ability to shut it off when you don't want it to!).

Wish you luck and if you have anything I can use to follow your work, would be interested!