Memory Out of Context: Spacing Effects and Decontextualization in a Computational Model of the Medial Temporal Lobe

Some neural representations gradually change across multiple timescales. Here we argue that modeling this "drift" could help explain the spacing effect (the long-term benefit of distributed learning), whereby differences between stored and current temporal context activity patterns produce greater error-driven learning. We trained a neurobiologically realistic model of the entorhinal cortex and hippocampus to learn paired associates alongside temporal context vectors that drifted between learning episodes and/or before final retention intervals. In line with spacing effects, greater drift led to better model recall after longer retention intervals. Dissecting model mechanisms revealed that greater drift increased error-driven learning, strengthened weights in slower drifting temporal context neurons (temporal abstraction), and improved direct cue-target associations (decontextualization). Intriguingly, these results suggest that decontextualization-generally ascribed only to the neocortex-can occur within the hippocampus itself. Altogether, our findings provide a mechanistic formalization for established learning concepts such as spacing effects and errors during learning.