It depends on the training policy you are using and also some parameters that are set in the training. I would suggest you reading this nice explanation written by Souvik about how the policies work and how the predictions are being made Trouble with 'There is no memorized next action' message.
To train your model you would definitely need at least a few sotires which would cover different conversation turns for your model to start making sensible predictions. The more good quality stories you will have, the better the predictions will be, just like in any other ML problem.