A common approach to prediction and planning in partially observable domains\nis to use recurrent neural networks (RNNs), which ideally develop and maintain\na latent memory about hidden, task-relevant factors. We hypothesize that many\nof these hidden factors in the physical world are constant over time, changing\nonly sparsely. To study this hypothesis, we propose Gated $L_0$ Regularized\nDynamics (GateL0RD), a novel recurrent architecture that incorporates the\ninductive bias to maintain stable, sparsely changing latent states. The bias is\nimplemented by means of a novel internal gating function and a penalty on the\n$L_0$ norm of latent state changes. We demonstrate that GateL0RD can compete\nwith or outperform state-of-the-art RNNs in a variety of partially observable\nprediction and control tasks. GateL0RD tends to encode the underlying\ngenerative factors of the environment, ignores spurious temporal dependencies,\nand generalizes better, improving sampling efficiency and overall performance\nin model-based planning and reinforcement learning tasks. Moreover, we show\nthat the developing latent states can be easily interpreted, which is a step\ntowards better explainability in RNNs.\n