MemoryGraph: uma proposta de memória para agentes conversacionais utilizando grafo de conhecimento

Abstract

With the advancement of massive language models applied to natural language processing, many proposals have become viable, such as the use of conversational agents applied to various everyday tasks. However, these models still have limitations in both the integration of new knowledge and the representation and retrieval of that knowledge, being constrained by costs, execution time, and training. Furthermore, their black-box nature prevents the direct manipulation of knowledge, mainly due to the vector representation that indirectly represents it, making the control and explanation of their results more difficult. In contrast, knowledge graphs allow for a rich and explicit representation of relationships between real-world entities. Despite the challenges in their construction, studies indicate that these can complement each other to produce better results. Therefore, the objective of this research is to propose a memory system for conversational agents based on massive language models through the combination of explicit knowledge (knowledge graphs) and implicit knowledge (language models) to achieve better semantic and lexical representation. This methodology was called MemoryGraph and is composed of three processes: graph construction, graph search, and user representation. Various knowledge graph construction workflows were proposed and compared, considering their costs and influences on the final result. The agent can search for information in this base through various search proposals based on RAG, referred to here as GraphRAG. This search methodology was evaluated by humans in five proposed question scenarios, showing superior average results in all five proposed search approaches (29% in the best approach). In addition, six RAG metrics, evaluated by a massive model, were applied to the proposed application results from two popular datasets and one composed of diabetes guidelines, showing superior results in all datasets. Furthermore, a method for long-term user representation, called user_memory, was proposed, demonstrating 93% retention of user information. To reinforce this result, case studies were conducted, demonstrating the agent's ability to personalize the user experience based on past experiences, increasing the speed of information delivery and user satisfaction. The results demonstrate that the MemoryGraph paradigm represents an advance over vector representation in environments where richer, temporal, and mutable contextualization is necessary. It also indicates that the integration of knowledge graphs with massive language models, especially in the construction of long-term memory and rich contextualization based on past experiences, can represent a significant advance in creating more efficient, personalized conversational agents with enhanced capacity for retaining and utilizing information over time.