'Gossip Cat ' is a project to build a cat avatar (running on a smartphone) which can gossip inconsequentially with the user.
My first thought was to model the content of gossip as a discussion about things. So Mrs Jones has a son, we have established, and the cat asks "Did Mrs Jones' son visit recently? " This is toggling an attribute value in a record structure.
But really gossip is not like this. It's more like a soap opera: all about agents - their situation, problems, goals, intentions, beliefs and knowledge in as transgressive a context as possible.
This might lead in the direction of baroque epistemic-conative logics but that would be a category error. Instead we should define structural models of agent situations and 'states of mind' which lend themselves to algorithmic manipulation.
I'm thinking of front-ending this with a robust, Eliza-like template-based natural language system.
A robust model of intentional dialogue, even if highly content-bounded (there's a reason it's a cat!) could lead to an interesting theory: I think this is the order to do it.