Hey @bndrq0!
Points 1 and 3 sound like what’s described in Contextual conversations.
Resetting the choices would then be equivalent to resetting the slot values, which can be done inside a custom action (see e.g. this forum thread), the action itself being triggered by some particular intent/entity.
“Self-learning”: There is no existing project for this. While it can be done, I also think that it’s too risky. How will the users know which of the intents their message falls into? Will the users always be honest? Will the users be able to specify not just the intent, but also which bot action should follow? This being said, you can tweak the NLU fallback flow to achieve something like what you’re describing.