The exponential forgetting function, which states that memory retention decreases as a negative exponential of the retention interval, represents the most elementary mathematical model of forgetting. Originally inspired by Ebbinghaus's (1885) pioneering studies of nonsense syllable retention, exponential decay has served as both a theoretical proposal and a null-hypothesis baseline in the study of memory dynamics.
The Exponential Function
The exponential forgetting function takes the form:
R = retention (proportion correct)
t = time since encoding
a = initial strength (R at t = 0)
λ = decay rate constant
The function has the property that the rate of forgetting at any time is proportional to the current strength: dR/dt = −λ · R(t). This means that a constant fraction of the remaining memory is lost per unit time, analogous to radioactive decay. The half-life of the memory is t₁/₂ = ln(2)/λ.
Exponential vs. Power Law
A central debate in the forgetting literature concerns whether forgetting is better described by an exponential function (R = a · e^(−λt)) or a power function (R = a · t^(−b)). Wixted and Ebbesen (1991) showed that the power function provides superior fits across a wide range of studies, particularly at long retention intervals where the exponential predicts too little retention. The power function's slower asymptotic approach to zero better matches the empirical observation that old memories are remarkably persistent.
Power: R(100) = a · 100^(−b) (much slower approach)
When Exponential Forgetting Is Appropriate
Despite the superiority of the power law in aggregate data, exponential forgetting may be appropriate for individual memory traces. Anderson and Tweney (1997) and Wixted (2004) argued that the power law in group data could emerge from averaging across individual traces that each decay exponentially but with different rate constants. This mixture hypothesis reconciles exponential mechanisms at the trace level with power-law behavior at the aggregate level:
Exponential Decay in Computational Models
Many computational models of working memory and short-term memory assume exponential decay as a default mechanism. The TBRS model assumes that unattended memory traces decay exponentially until refreshed. The ACT-R architecture uses a power-law decay for declarative memory chunks but exponential dynamics within individual activation computations. Neural network models often implement forgetting through exponential decay of activation or synaptic weights, as it corresponds to the simplest first-order differential equation: da/dt = −λa.
Ebbinghaus's original (1885) forgetting function was neither pure exponential nor pure power law, but a logarithmic function: R(t) = a / (log(t))^b + c. Modern reanalysis of his data shows that both power and exponential functions fit reasonably well over his relatively short retention intervals (minutes to days). The distinction between exponential and power functions becomes decisive only at longer time scales.