NETWORK ARCHITECTURE SHAPES MEMORY STORAGE IN HIPPOCAMPAL CA3 MODELS

How does network architecture shape the brain’s ability to store and recall memories? The hippocampal CA3 region—the brain’s largest auto-associative network—offers a natural testbed for this question. Here, we develop a recurrent CA3-like model that learns and recalls patterns to uncover the global and local circuit properties that maximize memory capacity.
We first asked which global architectural features best support memory storage. For this, we explored how to distribute a fixed total number of synapses across random excitatory networks of varying size and density—from small, highly connected networks to large, sparsely connected ones. Using analytical approximations that scale to realistic hippocampal dimensions, we found that distributing synapses across larger, more sparsely connected networks yields higher storage capacity than smaller, denser configurations. This relationship holds across synapse counts corresponding to mouse, rat, and human CA3, and is supported by Bayesian analyses showing that the same scaling emerges across a wide range of simulated parameters. These results suggest that evolution may have favored neuron number over connectivity, a trend matching anatomical data showing reduced CA3 connection probability from mouse to human.
We next examined how local structural motifs influence storage. Manipulating motif abundance revealed that non-random motifs among excitatory neurons enhance recall performance, consistent with experimental reports of CA3 microcircuit organization. Finally, we extended this framework to include inhibitory circuitry, exploring how excitatory–inhibitory structure modulates recall. Together, these results link hippocampal structure to function and provide testable predictions for how CA3 microcircuits support associative memory.