Episodic memories are encoded by sparse populations of neurons activated during an experience. These neural ensembles constitute memory engrams that are both necessary and sufficient for inducing recall even long after memory acquisition. This suggests that following encoding, engrams are stabilized to support reliable memory retrieval. However, little is known about the temporal evolution of engrams over the course of memory consolidation or how it impacts mnemonic properties. Here we employed computational and experimental approaches to examine how the composition and selectivity of engrams change with memory consolidation. We modeled engram cells using a biologically-plausible spiking recurrent neural network that yielded three testable predictions: memories transition from unselective to selective as neurons are removed from and added to the engram, blocking inhibitory neurons during recall disrupts memory selectivity, and blocking inhibitory synaptic plasticity during memory consolidation prevents engrams from becoming selective. By tagging activated neurons in vivo with high spatiotemporal precision as well as using optogenetic and chemogenetic techniques, we conducted contextual fear conditioning experiments that supported each of our model’s predictions. Our results reveal that engrams are dynamic even within hours of memory consolidation and that changes in engram composition mediated by inhibitory synaptic plasticity are crucial for the emergence of memory selectivity. These findings challenge classical theories of stable memory traces and point to a close link between engram state and memory expression.