According to the predictive coding framework, the brain automatically makes predictions about what is going to happen next. When predictions are incorrect, a prediction error (PE) arises, which is associated to memory updating. This may consist of encoding of the unexpected information as well as weakening of the original memory which the prediction was based on. However, it remains unclear when and how these similar memories can coexist.
In a five-session experiment, participants listened to naturalistic dialogues, which were encoded in an original form and later modified, inducing a PE. The modifications were either on the surface level, changing only the phrasing, or on a gist level, changing the overall meaning. Participants were then tested on original and modification recognition memory.
Both the originals and modifications were recognized best after higher-rated gist changes compared to smaller gist changes. For surface changes, the magnitude of modification had no effect. This is consistent with previous results and computational frameworks in that bigger PEs, induced by strong gist changes, lead to the encoding of a separate memory trace.
Using representational similarity analysis, we investigated whether original and modified dialogues are represented more distinctly (pattern separation) or similarly (memory integration). Results suggest that the two coding schemes are employed to differing extents, depending on PE size and quality. Whether these similar memories are integrated or separated can then have an impact on memory. Our results highlight that PEs can have several different outcomes and help to understand which circumstances lead to them.