In examining each recall transition, we next contrasted evidence for five strategies [temporal, semantic (via semantic analyses on the script), causal, chronological, and top causal (choosing the top causal scene)] (left). Next-scene recall was best explained by the top causal strategy (right). 6/

So, does time or another form of structure best explain recall organization?

We had participants watch and recall Memento, which separates time from causal/chronological structure. (Left: time vs. chronology; right: causal network structure, as normed by other participants). 4/

We found temporal abstraction by training the model w/ a range of ISIs (increasing powers of 2). Increasing ISIs resulted in greater strengthening in slower-drifting pools from EC-CA3. This helps retain memory access for longer according to the temporal regularity of training. 15/

Drift in this scenario (left) ⬆️ memory most at long RIs (middle), including w/ fully scrambled temporal contexts (final data point). Thus, drift ⬆️ memory even w/o contextual support. As predicted, drift ⬆️ error between model predictions and outcomes, as shown in hippocampal area CA3 (right). 14/

The relationship between ISI:RI is therefore non-monotonic, as exemplified by these findings (left) whereby the optimum ISI increases with increasing RI. Our model (right) captured these properties beautifully. (See the paper for other spacing effect simulations!) 12/

We used a biologically realistic model (below) of the entorhinal cortex and hippocampus -- two brain regions closely linked w/ memory. Temporal context, cue, and target patterns entered the model as separate “pools”. After training, we tested the model w/o targets to see if it reproduced them. 10/