Figure 4. The advantage of selectively encoding episodic memories at the end of an event.

(A) Prediction performance is better for models that selectively encode at the end of each event, compared to models that encode at the end of each event and also midway through each event. (B) The model performs worse with midway-encoded memories because midway-encoded target memories are activated more strongly than end-encoded target memories, thereby blocking recall of the (more informative) end-encoded target memories, and also because midway-encoded lure memories are more strongly activated than end-encoded lure memories (see text for additional discussion). (C) The cosine similarity between working memory states during part 2 and memories formed midway through part 1 (in orange) or at the end of part 1 (in blue). The result indicates that the midway-encoded memory will dominate the end-encoded memory for most time points. (D) The time-point-to-time-point cosine similarity matrix between working memory states from part 1 versus part 2 in the no memory (NM) condition (part C depicts the orange and blue rows from this matrix). (E) PCA plot of working memory states as a function of time, for a large number of events. The plot shows that differences in time within events are represented much more strongly than differences across events. The errorbars indicate 1SE across 15 models.