They use strictly disjoint sections of memory so using one does not reduce the space available for the other. Key moves are enough smaller that the keymove area of memory is much smaller than the learning area of memory but still has capacity for a larger number of signals.

You don't really care about "most efficient use of memory". You may care about fitting the largest possible number of functions. To do that, you would use both key moves and learned signals, selecting between them based on factors such as the fact that you can't use an RC5 function twice in a row if it's defined by learning (but for most other protocols you can).