This paper challenges the conventional view that volatility is smooth or has long memory by analyzing high-frequency financial data. The authors find that log-volatility behaves like a fractional Brownian motion with a Hurst exponent of approximately 0.1, indicating a rough, anti-persistent process. They propose the Rough Fractional Stochastic Volatility (RFSV) model, a variant of the Comte and Renault FSV model with H < 1/2, which captures this roughness. The RFSV model is shown to be remarkably consistent with empirical data and yields improved forecasts of realized volatility compared to standard models. Additionally, the paper demonstrates that statistical tests for long memory often incorrectly detect persistence in data generated by the rough process, explaining why long memory has been widely accepted as a stylized fact. Finally, the authors provide a market microstructure foundation, linking volatility roughness to high-frequency trading and order splitting, offering a quantitative explanation for the observed behavior.