The temporal activity of neural circuits exhibits fluctuations simultaneously varying over a large range of timescales (Murray et al. 2014). Recent experimental evidence presents a heterogeneity of timescales across neurons within the same local cortical circuit, ranging from milliseconds to minutes (Bernacchia et al. 2011). While this phenomenon is well documented, the underlying neural mechanism is still unknown. We present a novel random neural network whose units represent functional neural clusters of different sizes, given by a heterogeneous distribution of self-couplings. We find that the network activity varies over multiple timescales spanning several orders of magnitude. When the self-couplings are strong and heterogeneous, a reservoir of numerous timescales emerges in the chaotic phase, where a neural cluster’s timescale is proportional to its self-coupling. In the limit of large size, the dynamics of a network cluster can be studied analytically, revealing a new metastable regime described by colored noise-driven transitions between potential wells. In this regime, we provide a novel analytical treatment for recurrent neural networks, capturing their time-dependent metastable dynamics. When driving such a network with a time-dependent broadband input, given by a superposition of multiple frequencies, slow and fast neural clusters preferentially entrain slow and fast input frequencies, respectively, thus providing a potential mechanism for spectral demixing in cortical circuits. Our work establishes the basis of a novel framework for studying heterogeneity of timescales in neural circuits as well as artificial neural networks, highlighting the advantages of generating multiple timescales for encoding complex time-varying signals.