Techniques used in the area of pattern recognition, particularly the methods of unsupervised learning, are beginning to appear useful in the area of analyzing underwater acoustic signal spectra. A problem encountered in the use of these techniques for varied situations is that of an unlimited memory requirement when uncertain parameters are present in either the signal or noise process. The problem considered here is the design of a limited‐memory receiver for the unsupervised detection of signals in noise. Limited memory is achieved using quantization of a sufficient statistic, and the optimum‐likelihood receiver is developed under this constraint. The limited‐memory receiver is shown to retain adaptive or learning characteristics, and its performance is shown to converge to that of the unlimited‐memory receiver with increasing time. The primary intent of this work indicates the effect of increasing memory—in other words, increasingly finer quantization—on the performance of the limited‐memory receiver and compares this with the performance of the unlimited‐memory receiver.