While this might be a great idea, it would require in-depth testing to determine the impact on constantly varying video quality. The viewer will accept a degraded image that stays at the same amount of degradation throughout, but I don't think the experience will be as readily accepted if the video is high quality one moment and then changes to a lower quality the next.
Numerous codecs use a variable compression scheme, but they regularly write keyframes that contain a maximum of information, and then just write the changes in succeeding frames until enough change has occurred that another keyframe is required.
This is somewhat different than what you are suggesting since, in effect you would be changing the quality of the keyframes each time as well. (metaphorically speaking)
I'm not suggesting that the concept wouldn't work, just that it may result in a form of "image pumping" similar to the audio effect caused by an AGC circuit. This might be somewhat uncomfortable to watch.
I'd be interested in seeing the results if it is implemented.