Compress content to token budgets, prioritize context, summarize conversation history, built for LLM agents