Color constancy is the perceived stability of the color of objects despite changes in the light illuminating them. An object's color is considered constant if the current perceived color is judged to be in accord with the remembered one. Thus the accuracy and precision of color memory are fundamental to understanding this classic problem. Two hypotheses of color memory are tested here: (1) the photoreceptor hypothesis, which states that the color recalled from memory reproduces the light absorbed by each type of cone and (2) the surface-reflectance hypothesis, which states that the color recalled from memory is based on an inferred spectral reflectance of a surface that does not depend on the spectral distribution of the illuminant. In the experiments a test color is surrounded by either (i) a complex pattern composed of several colored patches or (ii) a uniform "gray" field at the chromaticity of the illuminant. In a control condition the test color is presented on a dark background. Long-term memory of the test color is measured in a production task begun 10 min after the end of the learning phase. In general, the results with a complex surround are consistent with the surface-reflectance hypothesis but not with the photoreceptor hypothesis. Color memory with the "gray" surround, on the other hand, shows a much stronger effect of the illuminant used during learning. These results are consistent with computational models of color constancy that require three or more chromaticities in view.