Large language models have a context window length limit, and when conversation history or tool call records become too long, the Agent needs to summarize and compress the history to free up space.