The point I was making is that whether or not something, especially a more complex algo like MTP, is memory-hard in practice can obviously depend on all sorts of details of the algorithm. And once such details have been spotted and fixed, I can't see how it would not be 'fundamentally' memory-hard.