recursive descent parsers generally work by doing something in the future based on the current token read, whereas Cognition works by getting a single token and then evaluating it or putting it on the stack based on the crank. If you think of syntax structure as if it is the AST itself then I can understand your confusion, but this is not the main point of the article anyways, and what I meant by this is simply that programming in some rigid way by which you read ahead based on the current token means you have something that's unchangeable in the interpreter design.
That said, I don't know if Cognition is unique. If you know of something that can do the same thing as this, be free to send that project our way.