In everyday experience, we encounter visual feature combinations. Some combinations are learned to support object recognition, and some are arbitrary and rapidly changing, so are retained briefly to complete ongoing tasks before being updated or forgotten. However, the boundary conditions between temporary retention of fleeting feature combinations and learning of feature bindings are unclear. Logie, Brockmole, and Vandenbroucke demonstrated that 60 repetitions of the same feature bindings for change detection resulted in no learning, but clear learning occurred with cued recall of the feature names. We extended those studies in two new experiments with the same array of colour-shape-location combinations repeated for 120 trials. In Experiment 1, change detection was well above chance from Trial 1, but improved only after 40 to 60 trials for participants who subsequently reported becoming aware of the repetition, and after 100 to 120 trials for participants reporting no awareness. Performance improved rapidly in Experiment 2 when participants reconstructed the array by selecting individual features from sets of colours, shapes, and locations. All participants subsequently reported becoming aware of the repetition. We conclude that change detection involves a visual cache memory that functions from the first trial, and retains feature bindings only for the duration of a trial. In addition, a weak residual episodic memory trace accumulates slowly across repetitions, eventually resulting in learning. Reconstructing feature combinations generates a much stronger episodic memory trace from trial to trial, and so learning is faster with performance supported both by the limited capacity visual cache and learning of the array.