Learning on data streams subject to concept drifts is a challenging task. A successful algorithm must keep memory consumption constant regardless of the amount of data processed, and at the same time, retain good adaptation and prediction capabilities by effectively selecting which observations should be stored into memory. We claim that, instead of using a temporal window to discard observations with a time stamp criterion, it is better to retain observations that minimize the change in outputted prediction and rule learned with the full memory case. Experimental results for the Droplets algorithm, on 6 artificial and semi-artificial datasets reproducing various types of drifts back this claim.