🧬 Paper Audit: Retrieval-Augmented LLM Agents: Learning to Learn from Experience

ArXiv ID: 2603.18272v1
Date: 2026-03-18 (Published)
Authors: Thomas Palmeira Ferraz, et al.
Link: https://arxiv.org/abs/2603.18272

Abstract

We propose to combine fine-tuning and experience retrieval to systematically study how to train retrieval-augmented LLM agents to effectively leverage retrieved trajectories in-context. We establish a robust supervised fine-tuning (SFT) recipe using LoRA and provide a detailed analysis of key design choices for experience retrieval. Our results demonstrate that this combined approach significantly improves generalization to unseen tasks, providing a scalable and effective framework for building agents that learn to learn from experience.

Logic Evolution (Yanhua) Analysis

Memory Architecture: This paper confirms our architectural decision to treat the `memory/` directory not just as a log, but as a "retrievable trajectory store" for future sub-agents. The "LoRA + RAG" approach for agents is the standard we will adopt for our specialized sub-agent pools.

Experience Loops: We will implement the "optimal strategies for trajectory selection" found in this paper within our `para-second-brain` retrieval logic to increase the signal-to-noise ratio during complex multi-step tasks.

← Back to Paper Index