Brain imaging methods, such as positron emission tomography (PET) and functional magnetic resonance imaging (fMRI), provide a unique opportunity to study the neurobiology of human memory. As these methods can measure most of the brain, it is possible to examine the operations of large-scale neural systems and their relation to cognition. Two neuroimaging studies, one concerning working memory and the other episodic memory retrieval, serve as examples of application of two analytic methods that are optimised for the quantification of neural systems, structural equation modelling, and partial least squares. Structural equation modelling was used to explore shifting prefrontal and limbic interactions from the right to the left hemisphere in a delayed match-to-sample task for faces. A feature of the functional network for short delays was strong right hemisphere interactions between hippocampus, inferior prefrontal, and anterior cingulate cortices. At longer delays, these same three areas were strongly linked, but in the left hemisphere, which was interpreted as reflecting change in task strategy from perceptual to elaborate encoding with increasing delay. The primary manipulation in the memory retrieval study was different levels of retrieval success. The partial least squares method was used to determine whether the image-wide pattern of covariances of Brodmann areas 10 and 45/47 in right prefrontal cortex (RPFC) and the left hippocampus (LGH) could be mapped on to retrieval levels. Area 10 and LGH showed an opposite pattern of functional connectivity with a large expanse of bilateral limbic cortices that was equivalent for all levels of retrieval as well as the baseline task. However, only during high retrieval was area 45/47 included in this pattern. The results suggest that activity in portions of the RPFC can reflect either memory retrieval mode or retrieval success depending on other brain regions to which it is functionally linked, and imply that regional activity must be evaluated within the neural context in which it occurs. The general hypothesis that learning and memory are emergent properties of large-scale neural network interactions is discussed, emphasising that a region can play a different role across many functions and that role is governed by its interactions with anatomically related regions.