GitTemporalAI: Leveraging Temporal Knowledge Graphs and LLMs for Multi-Agent Repository Intelligence

Dongsheng Luo, Raju Rangaswami, Amir Rahmati, Erez Zadok

Keywords: Multi-Agent Systems, Temporal Knowledge Graphs, Software Repository Mining

Abstract:

Large open-source software repositories represent complex multi-agent ecosystems where developers, code, and artifacts continuously interact and evolve. However, understanding these dynamic interactions and
leveraging them for automated issue resolution remains challenging. We present \emph{GitTemporalAI}, a system that constructs a temporal knowledge graph from repository data to capture the dynamic relationships between multiple entities (developers, files, issues, and pull requests) and their evolution over time. Our system employs a multi-agent architecture consisting of an embedding agent to encode repository entities, a search agent to traverse the temporal graph, and a reasoning agent that synthesizes contextual information to answer queries. By leveraging historical context and modeling relationships between repository entities, the system provides insights into repository evolution. Evaluation on a large open-source project, PyTorch Geometric, demonstrates GitTemporalAI's ability to improve query responses, particularly for tasks involving temporal reasoning and understanding repository dynamics.

Submission Number: 15