Why don't you look at the documentation on GPU programming?
Original scrypt was already compute bound. Neoscrypt is even more compute bound. People in GPU programming would go great lengths to have half the arithmetic intensity original scrypt had.
I'll ask you again: what does "memory intensive" mean? Fetching some memory every once in a while isn't an "intensive" operation.
My Cuckoo Cycle PoW spends 67% of its time on memory access. In other words, if you leave out all memory accesses, and just do the remaining computations, then runtime reduces to 33% of original.
That's what I would call "memory intensive".
You can do the same test for NeoScypt. Run it with all memory accesses removed, and see how much runtime is reduced.