In this article, we present a model of the representation of visual scenes in immediate memory. Our hypothesis is that the structure of this representation is equivalent to the construction of a Galois Lattice which it is based on principles of similarity and differentiation of objects through feature computation. The model allows making precise predictions about the effect of context, defined here as a more or less complex structure of features shared and not shared by an object to be memorized with other objects. We designed an immediate memory task of visually presented objects in which the number of objects and the number of properties remained constant. The distribution of these properties was manipulated. We hypothesized that for the same target object, errors rates as well as response times would prove to be a function of feature distribution. The results of the experimental study are consistent with our predictions and also allow reinterpreting the results of classic experiments in the field.