Macroeconomic time series often involve a threshold effect in their ARMA representation, and exhibit long memory features. In this paper we introduce a new class of threshold ARFIMA models to account for this. The threshold effect is introduced in the autoregressive and/or fractional integration parameters, and can be tested for using LM tests. Monte Carlo experiments show the desirable finite sample size and the power of the test with an exact maximum likelihood estimator of the long memory parameter. Simulations also show that a model selection strategy is available to discriminate between the competing threshold ARFIMA models. The methodology is applied to US unemployment rate data, where we find a significant threshold effect in the ARFIMA representation, and a better forecasting performance relative to TAR and symmetric ARFIMA models.