This paper investigates whether cued recall of multidimensional stimuli is all-or-none, as predicted by the Fragmentation Hypothesis (Jones, 1976); or probabilistic, as is commonly assumed in models of associative memory. To test this, composite stimuli were cued repeatedly, by each of their attributes in turn, to see whether the patterns of recall were consistent with all-or-none fragments. This test also requires a model to account for the inconsistent patterns of recall which are to be expected as a result of correct guessing. Of necessity, therefore, this paper also investigates the nature of guessing to enable the test of all-or-none recall.Two experiments are reported, both conditions of the same design and using the same pictorial stimuli. The first analysis looks at errors and seeks to discover interactions within sequences of responses. There are two stages involved. First, the most likely sources of each answer are indentified and classified according to a number of different categories of interest. Second, there is a statistical evaluation of the frequency with which these different categories occur. This analysis reveals: (a) the systematic recall of previous errors; and (b) that guesses comprising a pair of elements from the same (incorrect) stimulus occur more frequently than is expected by chance. Both processes have a systematic effect upon the pattern of correct guesses which is not predicted by the models of guessing commonly used.A model of cued recall is presented which combines the Fragmentation Hypothesis (including the assumption of all-or-none recall) with a model of guessing which hypothesises that a proportion of guesses act also as implicit cues for recall. This memory checking model of guessing is shown to predict the observed processes in guessing well. Overall, the combined model shows a satisfactory fit to the data, providing support for the all-or-none assertion. However, it is shown that a small proportion of inconsistent patterns of recall cannot be explained by the guessing model, and a low level of recall failure and forgetting is proposed. The recall of fragments is therefore closely approximated by all-or-none recall, but cannot be exactly so.