Being able to capture higher-order rules that govern sequences of, for example, robot actions is still a challenging task for neural networks. One of the mechanisms that allows to extract hierarchical dependencies between items within time series is rank-order coding. The different possible orders of items (actions), called ordinal codes, can be understood as a grammar, or a set of rules governing the sequences. We propose a sample-efficient neuro-symbolic architecture based around the mechanism of rank-order coding that is able to rapidly extract and learn the underlying rules from example sequences. These rules can then be used in a generative manner to construct novel sequences, adhering to the principle of compositionality (see Fig. 1). We demonstrate how this framework can be used with sequences of both abstract, compound elements, like discrete states of a game (Towers of Hanoi), as well as with sequences of actions performed in continuous motor space.