Well it depends on the model. Typically, online models forget older data in a transparent manner. For instance, a linear model forgets at a rate that is proportional to the learning rate that is used.