We constructed 4 working memory recognition models to predict behavior in the local recognition task (also called change detection), in which both content (e.g., color) and context (e.g., location) information are necessary to make correct recognition decisions. The theoretical assumptions incorporated in the models come from crossing 2 contrasts: One is the contrast between discrete-state models with continuous-strength models. The other contrast pertains to the dimensionality of information involved in the recognition process: either unidimensional (as in single-process recognition models) or two-dimensional (as in dual-process models). We compared the models to data from three local-recognition experiments using sequentially presented visual materials. All three experiments revealed intrusion costs (i.e., higher false alarms to probes matching a list element in the wrong context than to new probes) and U-shaped serial-position curves for all probe types. The two-dimensional continuous-strength model predicted these results best both qualitatively and quantitatively. The unidimensional and two-dimensional discrete-state models were able to predict the qualitative pattern of serial-position curves but failed to predict a sufficient amount of intrusion cost. The unidimensional continuous-strength model failed even to predict the qualitative pattern of the serial position effects.