A long-running research agent calls search and file-reading tools dozens of times per task. After many turns, the conversation nears the context window limit mostly because of old, bulky tool results the agent no longer needs. The team wants to keep recent work and the agent's conclusions while freeing space automatically. What should they do?