Concept learning is a form of supervised machine learning that operates on knowledge bases in description logics. State-of-the-art concept learners often rely on an iterative search through a countably infinite concept space. In each iteration, they retrieve instances of candidate solutions to select the best concept for the next iteration. While simple learning problems might require a few dozen concept retrieval calls to find a suitable solution, complex learning problems might require thousands of calls, making concept retrieval a major runtime bottleneck. We alleviate the resulting runtime challenge by presenting a semantics-aware caching approach. Our cache is essentially a subsumption-aware map that links concepts to a set of instances via crisp set operations. We evaluate our approach on four datasets using four symbolic reasoners, one neuro-symbolic reasoner, and five cache eviction policies. The results show that our cache reduces concept retrieval and concept learning runtimes by up to an order of magnitude while remaining effective across both symbolic and neuro-symbolic reasoning reasoners.