Why Empirical Forgetting Curves Deviate from Actual Forgetting Rates: A Distribution Model of Forgetting

For over a century, forgetting research has shown that recall decreases along a power or exponential function over time. It is tempting to assume that empirical forgetting curves are equivalent to the rate at which individual memories are forgotten. This assumption would be erroneous, because forgetting curves are influenced by an often-neglected factor: the distribution of memory strengths relative to a recall threshold. For example, if memories with normally distributed initial strengths were forgotten at a linear rate, percent correct would not be linear, it would decrease rapidly when the peak of the distribution was crossing the recall threshold and slowly when one of the tails was crossing the threshold. We describe a distribution model of memory that explains the divergence between forgetting curves and item forgetting rates. The model predicts that forgetting curves can be approximately linear (or even concave, like the right side of a frown) when percent correct is high. This prediction was supported by previous evidence and an experiment where participants learned word pairs to a criterion. Beyond its theoretical implications, the distribution model also has implications for education: Creating memories that are just above the threshold helps on short-term tests but does not form lasting memories.