Sequential analyses of learning data collected in noncontingent 2-choice probability learning studies often reveal that a response made on trial n is strongly influenced by the pattern of events which occurred over the last few trials. A learning model which incorporates such a memory of recent events as a basic assumption has been found recently to be only approximately correct. The present study introduces a new memory test which attempts to assess some of the assumptions made by the k-span model. The data suggest that certain modifications of the model are called for.