The model state encodes the past tokens (in some lossy way that the model has chosen for itself). You can ask it to try and, assuming its attention is well-trained, it will probably do a pretty good job. Being able to refer to what is in its context window is an important part of being able to predict the next token, after all.