From John Locke until the present day, temporal pairing has been assumed to be the sine qua non for the formation of an association. Association formation is at the heart of connectionist modeling and of efforts by neurobiologists to discover the physical basis of memory. However, temporal pairing has never been objectively defined for any experimental paradigm. Moreover, under any definition ever proffered, it has been shown to be neither necessary nor sufficient for producing conditioned behavior in associative learning paradigms. I argue that what really drives conditioning is that component of the mutual information between a predictive stimulus (CS in the jargon of the field) and the predicted stimulus (US) that depends on the protocol parameters. This assumption permits me to derive from first principles the observed empirical effects of varying the temporal parameters in conditioning protocols. What modelers and neurobiologists should be focusing on is mutual information and mechanisms for computing it, not temporal pairing.