We present a dynamic model that explains the recurrence of tax amnesties as a self-reinforcing phenomenon driven by government reputation. We model repeated interactions between a government and a continuum of taxpayers, where the government’s type evolves stochastically and remains unobservable to taxpayers. Tax amnesties serve as signals of the government’s optimizing behavior, influence taxpayers’ expectations of future amnesties, and incentivize strategic tax evasion. This expectation-driven evasion behavior reduces immediate tax revenues, increases the fiscal attractiveness of subsequent amnesties, and creates a “reputation trap” that perpetuates amnesty cycles. We show that the mechanism is robust across all Markov-perfect equilibria, with the amnesty probability peaking after prior amnesties. The model accounts for variation in tax amnesty frequency across U.S. states and shows how factors such as income, tax rates, audit intensity, and amnesty costs affect vulnerability to reputation traps. Our findings highlight the role of reputation as an endogenous commitment device in dynamic fiscal policy.