Seems like an intriguing product, though certainly with the potential to be "creepy" if executed in the wrong way. I was struck by an implementation paragraph though:<p>> Every major component of Refresh is a fuzzy system. When it sees “Lunch with Tom” – without exaggeration 100 bots go out and vote who Tom is. One says “You have a Facebook friend named Tom”. One says “You met with a Tom person 3 months ago at the same location.” One says “It’s not your cousin Tom — you hate him.” Some are heuristic and some are statistical.<p>Putting my grad-student hat on, from a Bayesian perspective, they seem to be effectively "marginalizing" the distribution on "facts" over the hidden identity-of-Tom variable given the event and your current network:<p>$ p(facts | network, event) = \int p(facts | identity) p(identity | network, event) d identity $<p>The app's output seems to be, effectively, samples from that "facts" distribution, which can be generated by taking samples from the identity distribution and accepting/rejecting based on likelihood. It's a more formal way to look at fuzzy logic systems like this one. The cool thing is that you can eventually plug other things into the system, such as taking into account the distribution over types of facts that the current user likes to look at, in a mathematically sound way. But it's rare (imho) to come across systems like this outside of academia.