^As well as that, the old school solution was to throw out the Z channel of your normal maps (because you can reconstruct it from the other two), and then putting the X/Y channels into the Green and Alpha channels of a DXT5 (with nothing in Red/Blue).
This gives better quality than naively using RGB, because DXT5 compresses RGB and Alpha data seperately, so by using G and A, you make sure there's no cross-talk during compression. The alternative to achieve the same memory savings is to reduce the resolution by half in both width and height (1/4 the pixels).

If you're targeting hardware without much memory, then compression or low resolution is a requirement. We used config files in our build system to swap between the two to compare quality, and also we let the artists specify a choice per asset if required.