If what Mark said is true, and the frames are simply an internal data structure within the kernel (and not their own separate kernel space), then the re-using of frames keeps the kernel from taking up too much flash memory.

They originally intended to actually delta-compress the frames using a Real Man's Algorithm, but ran out of steam somewhere along the path to 2.0...

I seem to recall Rob lamenting that there are still still some unresolved issues with the amount of flash they take up even now.