How would you design a prompt to handle multi-turn dialogue with memory constraints?

• A. Use concise turn-taking prompts, maintain a short-term memory with summarized state, and ask for user confirmation to avoid drift; optionally store long-term memory externally.
• B. Use extremely long prompts that store everything.
• C. Avoid any memory management to keep prompts simple.
• D. Only rely on the model's internal memory without external storage.

Answer: (A)

Handling multi-turn dialogue under memory limits means designing prompts that keep context compact, track essential state, and guard against drift. Using concise turn-taking prompts helps by keeping each user and assistant turn small, so the model can respond effectively without being overwhelmed by a long prompt. Building a short-term memory with a summarized state captures the important aspects of the conversation—like user goals, constraints, and key decisions—without repeating everything, which preserves coherence while staying within token limits. Asking for user confirmation at points of potential ambiguity helps prevent drift, ensuring decisions stay aligned with user intent as the dialogue progresses. Storing long-term memory externally gives you a place to keep persistent preferences, facts, or task-specific knowledge and retrieve them when needed, rather than trying to cram everything into the current prompt.

The other approaches run into practical problems: stuffing everything into one long prompt isn’t scalable due to token limits and cost, leaving memory management out means the system can lose track of prior turns and user goals, and relying solely on the model’s internal memory without external storage can lead to forgotten context across long conversations or between sessions.