In this work we reevaluate and elaborate Crick-Mitchison's proposal that REM-sleep corresponds to a self-organized process for unlearning attractors in neural networks. This reformulation is made at the face of recent findings concerning the intense activation of the amygdalar complex during REM-sleep, the role of endocannabinoids in synaptic weakening and neural network models with itinerant associative dynamics. We distinguish between a neurological REM-sleep function and a related evolutionary/behavioral dreaming function. At the neurological level, we propose that REM-sleep regulates excessive plasticity and weakens over stable brain activation patterns, specially in the amygdala, hippocampus and motor systems. At the behavioral level, we propose that dream narrative evolved as exploratory behavior made in a virtual environment promoting "emotional (un)learning", that is, habituation of emotional responses, anxiety and fear. We make several experimental predictions at variance with those of Memory Consolidation Hipothesis. We also predict that the "replay" of cells ensembles is done at an increasing faster pace along REM-sleep.