Compress long AI-agent context into a concise summary, facts, or episodic memory while preserving key entities.