Crucially, we show that the more precisely hippocampal spiking variability tracks the composite (late-layer) features that comprise each individual stimulus, the better those stimuli are later remembered up to 30 minutes later. These effects were also *spatially specific* to the hippocampus.

Hippocampus spiking entropy was coupled to image features in every.single.subject...with tighter coupling to more composite (late layer) than to simpler (early layer) image features.

We used comp. vision models (HMAX & VGG16) to estimate the features of images presented to patients (N = 34) during memory encoding. We then estimated the coupling of hippocampal single-neuron variability (entropy) to image features via within-participant latent correlations.

We argue that the architecture of multi-layer computational vision models can be used to differentiate between simple and composite visual features of any stimulus a participant may encode.