Memory interface can be thought of as the "highway" that videocard data travels on.
A 128-lane highway can move twice as many cars as a 64-lane highway, assuming the cars are travelling at the same speed, say, 60 mph.
However, a 64-lane highway with cars travelling at 120 mph is moving the same number of cars a 128-lane highway with cars travelling at 60 mph can move, dig?
In the same way, a Geforce 6600 with a 128-bit memory interface and a 600 mhz memory speed is moving half of the data a Geforce 6800 with a 256-bit memory interface @ 600 Mhz will move.
However, memory running at 1200 Mhz on a 128-bit interface will do comperable work to 600 mhz memory on a 256-bit memory interface.
This is why it's important to avoid "le" versions of videocards with 64-bit memory interfaces.
Memory interface and memory clockspeed are important to videocard performance, but the memory interface has a larger impact in most cases because inferior interfaces can quickly halve memory performance.
As far as pipelines vs. Memory interface, well, they're both very important. More Pipelines generally means more GPU processing power, and more memory bandwidth generally means better performance with AA/AF and at higher resolutions. So it's hand and hand.