HighlightsRole of prediction errors (PE) in human one-shot declarative learning.PE is manipulated via previous experiences (priors) and sensory inputs (evidence).PE leads to superior memory across 5 different experiments.Support for Predictive Interactive Multiple Memory Signals (PIMMS).

Previous work has shown that priming improves subsequent episodic memory, i.e, memory for the context in which an item is presented is improved if that item has been seen previously. We previously attributed this effect of "Priming on Subsequent Episodic Memory" (PSEM) to a sharpening of the perceptual/conceptual representation of an item, which improves its associability with an (arbitrary) background context, by virtue of increasing prediction error (Greve et al, 2017). However, an alternative explanation is that priming reduces the attentional resources needed to process an item, leaving more residual resources to encode its context. We report four experiments that tested this alternative, resource-based hypothesis, based on the assumption that reducing the available attentional resources by a concurrent load would reduce the size of the PSEM. In no experiment did we find a significant interaction between attentional load and priming on mean memory performance, and nor did we find a consistent correlation across participants between priming and PSEM, failing to support the resource account. However, formal modelling revealed that a resource account is not, in fact, inconsistent with our data, by confirming that nonlinear (sigmoidal) resource-performance functions can reproduce any interaction with load, and, more strikingly, any pattern of correlation between priming and PSEM. This work reinforces not only the difficulty of refuting attentional resource accounts of memory encoding, but also questions the value of load manipulations more generally.