Generally each card renders a full alternate frame, and thus each needs its own memory resources to draw from, the frames are then interleaved splitting the load across 2 cards.
If they were to share ram, the data rate and latency would get greatly sacrificed over the pcie bus compared to the direct connection the gpu has with it's own on-board ram, not even sure how it would work with AFR. Not only the latency but you essentially half the bandwidth with just having 2 gpu's trying to acess that. This may change in the future with dx12 where it can be shared because the gpu resources are all working on one frame together, but not sure how well that will work. It certainly wont get the same performance as each card with its own ram.