MemPrompt: Memory-assisted Prompt Editing with User Feedback

TLDR: A method to "fix" GPT-3 After Deployment with user interaction.

MemPrompt maintains a memory of errors and user feedback, and uses them to prevent repetition of mistakes.

Blog, Code, and Paper

arxiv arxiv github

Three minute overview

MemPrompt Overview

MemPrompt: Workflow

Adding New Feedback To Memory

Memory is populated with new feedback: A user enters a question for which no feedback is available (steps 1, 2). Directly prompting GPT-3 with the question leads to incorrect answer and understanding (step 3). User-provides feedback on the incorrect understanding (step 4), which is added to memory (step 5).

Utilizing Old Feedback To Answer New Questions

Retrieving feedback from memory: A user enters a question which GPT-3 has incorrectly answered in the past, and has received feedback from a user (step 1). The feedback is retrieved from memory (step 2), and both question and feedback are added to the prompt. The prompt contains examples that allow GPT-3 to react to user feedback and generate correct understanding and answer.

More resources

Checkout the videos below from Yannic Kilcher's Machine Learning Youtube channel MemPrompt.
Overview of the method
Discussion with the Authors