I think it is an intriguing idea. I was just speaking with some EMC people about this sort of thing. In benchmarking their SSD / Flash drives they were able to clock the disks up to 50,000 IOPS (but classify them at 2500 IOPS). Take this next part with a grain of salt - They said the disks performance gains stopped at 30,000 due to the bottleneck now being in the controller cache.

If I could get 30,000 IOPS out of the Drives, I wouldn't bother with an in-memory database. With disks becoming so much faster, would there be any point in shelling out the research for this type of database? I think I would rather see other things developed first.