Semantic caching stores responses keyed by meaning, not exact text, so paraphrased agent queries hit the cache and skip a re-call to the model. Here's how to implement it, tune the similarity threshold, and design for the failure modes.