The brain continuously makes predictions about what is going to happen next and uses episodic memories in the process. If there is a mismatch, a prediction error arises and can lead to memory updating, consisting of new learning, and forgetting of the old episode in order to improve future predictions.

We tested whether these effects extend to naturalistic conversations and whether the size and quality of prediction errors influence this process. Participants encoded original versions of short conversations and later heard the same dialogue in a modification. We manipulated encoding frequency of the original episodes (prior precision) and amount of difference between originals and modifications (prior accuracy). Furthermore, some changes consisted merely of paraphrases (surface level), while others altered the meaning of the statement (gist level). In a recognition test, we tested memory for the original as well as the modification.

While changes in the dialogues generally led to reduced recognition of the original version, this effect was slightly more pronounced with medium prediction errors, produced by big changes with weak priors or small changes with precise priors. Furthermore, bigger gist changes led to increased recognition of modifications, while surface changes did not trigger new encoding. Our findings indicate that effects of new learning in response to prediction errors extend to naturalistic conversations. They also highlight that not only the size of the prediction error, but also its quality influence memory outcomes, with new learning primarily occurring when critical new information is offered through gist changes.