When we remember an experience, patterns of neural activity present during encoding can reappear during retrieval, a phenomenon known as neural reinstatement. During silent retrieval, however, the content being remembered at each moment is unknown, making it unclear what temporal unit should be used to compare encoding and retrieval activity. We compared three approaches in an existing fMRI dataset from 32 participants who viewed six aversive film clips and silently retrieved three of them approximately 24 hours later: activity averaged across the full encoding and retrieval windows, individual fMRI time points (TRs), and data-driven neural states identified using Greedy State Boundary Search (GSBS). Film-specific encoding–retrieval similarity (ERS) was calculated as similarity to the matching encoded film minus the mean similarity to the five nonmatching films. None of the three approaches produced parcel-level or cortex-wide reinstatement effects that survived false-discovery-rate correction across the 401 spatial analyses.

However, cortex-wide ERS differed overall across methods, with the largest mean ERS for whole-window averaging (M = .041), compared with TR×TR (M = .003) and GSBS (M = .001), although individual pairwise comparisons did not survive correction. Exploratory analyses showed no systematic increase in ERS as local temporal bins became progressively coarser, and GSBS did not outperform fixed bins matched on window duration and number of segments. Film identity also explained substantial variability in ERS. Overall, preserving finer temporal structure did not improve film-specific reinstatement detection in this dataset, and data-driven neural-state boundaries provided no detectable advantage over matched regular segmentation.