The selected design maps to It prevents the agent from forgetting past successes and failures, which is the highest-control path for this scenario rather than a prompt-only or single-service shortcut. For knowledge-grounded agents, the clean architecture is a RAG path with retrievers and vector indexes externalized from the LLM, then evaluated for retrieval quality and answer faithfulness. Agentic systems need explicit decomposition: a planner or coordinator defines the work, specialized agents or tools execute bounded actions, and memory/state is preserved only where it improves the next decision. That structure increases maintainability because each agent role, message contract, and state transition can be tested independently under load. The distractors are weaker because they lean on A: It prevents the agent from becoming overly specialized to a single task; B: It eliminates the need for external storage like RAG; D: It guarantees the agent will produce the same output for the same..., which compromises traceability, resilience, scalability, or policy enforcement in production. The answer therefore fits NVIDIA's production-agent pattern: modular workflow design, measurable runtime behavior, GPU-aware serving where applicable, and controlled integration with enterprise systems.