Prompt caching is a speed and cost helper for prompts that keep starting with the same long instructions or documents. It isn’t memory, which brings a saved fact forward. It’s mechanical: the system notices a large chunk of the prompt is the same as before and reuses the work it already did. It also doesn’t make the context window bigger, since the model still has a limit on what it can see at once.
Think of a thick client binder on your desk. If a coworker asks ten questions about the same client, you don’t reread the whole binder from page one before each answer. You read it once, keep it open, and use it while answering.
How it shows up
AI tools often carry a lot of repeated setup: a system prompt, project instructions, glossary rules, long style guides, included again and again, all costing tokens to read. You usually won’t manage this by hand in client work; it’s more common in API systems and agent frameworks, where a developer puts the stable part first. The stable part is the binder, the changing part is today’s question. It connects to a normal cache, which keeps recently used work close by so the system doesn’t redo expensive work every time.
Why you care
Repeated context has a cost, and stable, well-organized instructions are easier for both people and systems to reuse. Keep rewriting the whole prompt in a slightly different order and the system can reuse less. If your AI setup repeats the same long material every day, prompt caching is one way it avoids paying the full price every time.