AWS has published a guide describing prompt-caching patterns for AI agents that use OpenAI models on Amazon Bedrock. The post says agents typically resend tool definitions, instructions, prior tool calls and results with each request because the model does not retain internal memory between requests.

According to AWS, prompt caching stores the processed state of a reusable input prefix and loads it for later requests. This can reduce the amount of repeated input processing and, the company says, lower input costs and latency. The post describes two approaches: implicit caching, in which Amazon Bedrock selects a reusable prefix as a conversation grows, and explicit caching, in which developers mark the end of content to save with a prompt-cache breakpoint.

AWS recommends combining the approaches when an agent needs both a reusable starting point and cached conversation history. The examples place stable tools and instructions before changing content such as user requests, dates and request IDs. The post also says that cached text must remain exactly the same, including the serialization of instructions and tool lists.

The guide covers continuing conversations, caching tool results and reusing shared content when agents start new sessions or spawn helper agents. It gives an example in which a roughly 4,000-token tools-and-instructions prefix writes 4,120 tokens on the first request and reads those tokens on later requests. AWS says a reusable prefix must reach 1,024 tokens to qualify, and the examples use a 30-minute cache time-to-live.