I think you're doing what Eliezer_Yudkowsky just warned against: how long is that hypothesis?

"The brain 'decides' to erase dream memories in such a way that we lose ~80% of our short-term memory of them and all of the long-term memory of them that we don't deliberately try to remember, on the basis of sufficient Bayes-structure in the brain, and its having observed (over period _?) the lack of entanglement between dream-based conclusions and the real world, and its prediction of sufficiently bad consequences ..."

Also, isn't that brain architecture a pretty narrow target for such a dumb process as evolution to hit? It's smart enough to exterminate dream conclusions (not to mention identify what is a dream, before you wake up) with extreme prejudice but not e.g. astrology?

FWIW, here's the theory I prefer and would defend:

"Dreams are random neuron firings" + "The feeling of dream coherency is the result of the brain's normal procedure for mapping sense data to best hypothesis."