MemesOverTime: A Decade-Scale Study of Meme Dynamics
Abstract
AbstractInternet memes evolve as they are repurposed across online communities, yet studying how their usage and language shift requires corpora that follow the same templates over many years, which existing resources rarely provide. To address this gap, we collect a decade of posts from Reddit meme subreddits (2016–2025) and introduce MemesOverTime, a semi-automatically curated dataset of 6.6M memes annotated with 6,584 unique templates. We use it to conduct an empirical study of meme template dynamics. We find that templates differ markedly in their temporal activity patterns: most reach peak usage within the first third of their active period, with short-lived templates peaking especially early. For long-lived templates, lexical drift is generally modest, with templates with more instances showing greater stability. Case studies reveal that creators spontaneously develop conventionalized expressions for some templates, while topical references shift over time.