What about going logarithmic with time? The more it's in the past the more errors are forgiven. Explained in another way use the relative time difference (diff/(now-value)) to compute the score and not the absolute time difference