Innovative Memory Retrieval with Prospection Techniques

In the realm of personalized dialogue systems, effectively retrieving user-specific information from extensive interaction histories is crucial. Traditional methods, such as Retrieval-Augmented Generation (RAG) and GraphRAG, typically focus on the immediate semantic similarity between queries and stored facts. This often leads to missed opportunities for relevant information that may not closely align with the original query. To address this limitation, we propose a new method called Prospection-Guided Retrieval (PGR). This approach separates the retrieval process from memory storage by utilizing a user's goals to generate a Tree-of-Thought (ToT) or a sequence of likely next actions. These generated steps serve as retrieval probes, enhancing the chances of uncovering pertinent facts that are not directly linked to the initial query. The retrieved information then informs the next phase of prospection, allowing for the discovery of additional relevant memories. We also introduce MemoryQuest, a rigorous benchmark that evaluates this method across various user profiles. Results indicate that PGR significantly enhances retrieval performance, achieving nearly three times the recall rate compared to existing methods. Furthermore, user evaluations show a strong preference for PGR-generated responses, highlighting its effectiveness in improving long-term retrieval and response quality.