Using retrieval contingencies to understand memory integration and inference

Past work has yielded mixed insights into how people draw upon their memories to make flexible new inferences. Neuroimaging approaches have shown that memories can be combined during encoding to store never-experienced, inferential associations. By contrast, behavioural research has emphasized a retrieval-based mechanism in which separate, high-quality memories are recombined as inferences are needed. We hypothesized that there might be important individual differences to consider when reconciling these seemingly disparate findings. We set out to quantify these differences by measuring contingencies in people’s memory recall behaviour. In Experiment 1, we first compared the performance of three memory contingency metrics using simulations and data from a task known to induce dependency. In doing so, we developed a correction to remove biases associated with general memory performance to isolate the representational structure of memories, and we selected the highest-fidelity option—corrected Dependency—for subsequent analyses. Experiment 2 tested the sensitivity of our chosen metric: We manipulated the similarity across experiences to encourage integration for half of the memories. Consistent with prior work, we found that increasing the similarity between the two experiences yielded reliable recall dependency. Finally, in Experiment 3, we used memory dependencies to reveal individual differences in inference approaches: While indeed “separators” relied upon high-fidelity individual memories to make inferences, “integrators” made inferences just as accurately and more rapidly than separators, regardless of how well they recalled constituent experience details. Together, these findings highlight the importance of considering individual differences in memory representations when characterizing the mechanisms underlying flexible inference.