Each layer of the network is like a step, and each token prediction is a repeat of those layers with the previous output fed back into it. So you have steps and a memory.