“Similarity” is often thought to dictate memory errors. For example, in visual memory, memory judgements of lures are related to their psychophysical similarity to targets: an approximately exponential function in stimulus space (Schurgin et al. 2020). However, similarity is ill-defined for more complex stimuli, and memory errors seem to depend on all the remembered items, not just pairwise similarity. Such effects can be captured by a model that views similarity as a byproduct of Bayesian generalization (Tenenbaum & Griffiths, 2001). Here we ask whether the propensity of people to generalize from a set to an item predicts memory errors to that item. We use the “number game” generalization task to collect human judgements about set membership for symbolic numbers and show that memory errors for numbers are consistent with these generalization judgements rather than pairwise similarity. These results suggest that generalization propensity, rather than “similarity”, drives memory errors.

Whether estimating the size of a crowd, thinking about how heavy something will be before trying to lift it, or reviewing a restaurant on a five-star scale, the modern world frequently requires us to navigate between subjective sensory experiences and shared formal systems. This entails mapping internal variables onto a common scale. Here we ask how people manage this in the case of estimating number. We present people with arrays of dots and ask them to report how many dots there are. In Experiment 1, we test predictions made by existing models of how people map from internal representations of numerosity to verbal estimates. We find that people’s estimates do not have a stable coefficient of variation at higher magnitudes, as has previously been suggested, and that the likely cause of this is a “drift” in people’s estimate calibration over many trials. Building on these results, we present a model of the mapping function from subjective numerosity to formal number estimates which relies only on a limited set of previous estimates and a rough sensitivity to the distribution of numbers in the world. Our model is able to generate an accurate mapping with limited data, as well as reproduce the notable aspects of human estimation described in our experimental results, namely humanlike patterns of underestimation, individual variability, and dynamic miscalibration at higher magnitudes.