Figure 3. The learned episodic retrieval policy is selective.

Panels A, B and C show the model’s behavioral performance, memory activation, and episodic memory gate (EM gate) value during part 2, across the recent memory (RM), distant memory (DM), and no memory (NM) conditions, when the penalty for incorrect prediction is set to two at test. These results show that recall is much stronger in the DM condition (where episodic retrieval is needed to fill in gaps and resolve uncertainty) compared to the RM condition. (D) shows that, in the DM condition, the EM gate value is lower if the model has recently (i.e., in the current event) observed the feature that controls the upcoming state transition. (E) shows how the average recall time is delayed when the penalty for making incorrect predictions is higher. (F) illustrates the definition of the schema strength for a given time point. (G) shows how the average EM gate value changes as a function of schema strength (penalty level = 2). The errorbars indicate 1SE across 15 models.