Better understanding of the Atlantic Multidecadal Oscillation (AMO) temporal evolution and the precise and reliable predictability of its phase-switching behaviour are of utmost importance in characterizing future climate change as well as its consequences. For the first time, this study investigated internal dynamics of the AMO in different warm/cold phases to examine whether this phenomenon is governed by a non-linear process through time or not. Also, changes in heteroscedasticity of the AMO and its memory structure were explored using Generalized AutoRegressive Conditional Heteroscedasticity (GARCH) models. The developed GARCH models showed a high volatility or a rapid change of the heteroscedasticity of the AMO and its different phases while the memory in the variance regarding the conditional variance parameter was not significant except the first and last warm phases. Also, the volatility parameter showed very high increasing rate from 1875 to 2018. This implies that the fluctuation in the variance has been increasing and the memory in the AMO has been declining in recent decades. The test for stationarity revealed the tendency of increasing volatility in recent decades. The change in the conditional variance is exponentially increasing by the AMO values and the non-linearity has been clearly increasing in recent decades. This non-stationarity, volatility and nonlinearity surge may result in less predictability of the AMO behaviour/phase-switching pattern as well as less predictability of its consequences and effects on the global atmospheric and oceanic processes in future.