Distilling Feedback into Memory-as-a-Tool

Victor Gallego

Keywords: file-based memory, tool-calling, self-critique

TL;DR: The paper proposes "Memory-as-a-Tool," a framework that reduces the high cost of iterative LLM self-correction by distilling feedback into persistent, human-readable guidelines through a file-based memory

Abstract:

We propose a framework that amortizes the cost of inference-time reasoning by converting transient critiques into retrievable guidelines, through a file-based memory system and agent-controlled tool calls. We evaluate this method on the Rubric Feedback Bench, a novel dataset for rubric-based learning. Experiments demonstrate that our augmented LLMs rapidly match the performance of test-time refinement pipelines while drastically reducing inference cost.

Code: https://github.com/vicgalle/feedback-memory-as-a-tool
Data: https://huggingface.co/vicgalle/rubric-feedback-bench

Submission Number: 1