My hypothesis is that all RNNs (and in general complex dynamical systems) need to be reset periodically. If run for too long without resetting, they tend to get stuck in strange states, blow up, or cease activating. You can see this effect by running a generative RNN model for a long time - eventually the output is garbage.

Under this model the next obvious question is why it takes so long to reset the brain's state. Maybe it can be done faster.