Paired associative inference (PAI) task and schematic of difference between encoding-based and retrieval-based models of inference. (A) Overall design of PAI task: participants see overlapping object pairs during study (e.g., AB, BC) and are required to generalize during the test phase by exploiting indirect relationships between objects (e.g., A—C): see main text for details. (B) Schematic of difference at the representational level between encoding-based and retrieval-based models, exemplified by the relational theory and REMERGE model, respectively. Left panel: encoding-based models (e.g., relational theory: Cohen and Eichenbaum): A—C relationship represented through overlapping conjunctive representations in the hippocampus which are created during the study phase. In this highly simplified illustration, each circle denotes a single neuron (or population of neurons)—red circles denote neurons which code for study episode AB and yellow circles denote those coding for episode BC. Orange circles denote neurons that participate in the representation of both AB and BC episodes (cf. nodal codings—see main text). Neuronal representations of AB and BC are overlapping, as if within a “memory space” (see Eichenbaum et al., 1999). Right panel: Retrieval-based models (e.g., REMERGE: Kumaran and McClelland) emphasize a principle of pattern separation in the hippocampus—depicted in extreme form as zero overlap between related episodes (AB, BC). Indirect relationships are not actually represented, therefore, but computed at the point of retrieval—which in REMERGE occurs through a recurrent mechanism which supports the dynamic interaction of pattern separated codes for related experiences.