Retrieval-Augmented Generation (RAG) is a promising approach that integrates nonparametric knowledge from external knowledge bases into models to improve response accuracy and mitigate factual errors and hallucinations. However, existing RAG methods perform independent retrieval tasks and directly integrate the retrieved information into the generation process without maintaining a summary memory or using an adaptive retrieval strategy. Consequently, they struggle with open-domain QA tasks due to noise caused by redundant information and insufficient information integration. In this paper, we propose Adaptive Memory-Based Optimization (Amber) for an improved RAG for open-domain QA tasks. Amber consists of an agent-based memory updater, an adaptive information collector, and a multi-grain content filter, all operating in an iterative memory update paradigm. A multi-agent collaborative approach integrates and optimizes the memory of the language model, ensuring comprehensive knowledge integration from previous retrieval steps. It dynamically adjusts the retrieval query based on accumulated knowledge and determines when to stop the retrieval, thereby enhancing retrieval efficiency and effectiveness. Furthermore, it filters irrelevant content at multiple levels to reduce noise and retain essential information, improving overall model performance. We demonstrate the superiority and effectiveness of our method and components through extensive experiments on several open-domain QA datasets.
