Other plausible solutions

It seems to me that the modified fatigue model is right but incomplete.
In our experiments, learning was turned off.
This may be reasonable especially in the circumstance when the concept is well known.
However, some short-term potentiation, may allow the circuit to persist longer when reactivated.
Another option is to change the topology.
This is not wiring connections, but allowing the topology to direct CA formation so that some neurons are more or less involved in the CA. When the CA is reactivated, new less involved neurons begin to participate.
A third option is to have different classes of neurons (as appear in the brain) with one set of neurons or neuron pairs responsible for CA activity levels.
Of course these solutions are not necessarily independent.