Memory-assisted prompt editing after deployment
EMNLP 2022: grow a memory of misunderstandings plus user clarifications, then edit prompts for similar future queries—no retraining.
Source: Memory-assisted prompt editing to improve GPT-3 after deployment
Madaan et al. treat post-deployment mistakes as first-class data. Pair the model with a growing memory of cases where it misunderstood intent, plus the user’s clarification, then retrieve similar cases to edit prompts for new queries—without retraining GPT-3.
It is an early, clean statement of the feedback-memory loop: record the correction, retrieve it, change behavior. Simulated-user experiments show large accuracy lifts on queries that used to trigger systematic misunderstandings.
Why it matters for code review
A good review comment is a clarification of intent or contract. If you store those clarifications and retrieve them on similar diffs, you get a lightweight “never make that mistake twice” system without waiting for fine-tunes. This paper is the pre-agentic ancestor of that pattern.