Improving OpenDevin: Boosting code generation LLM through advanced memory management

OpenDevin, a code generation AI tool, has emerged as a powerful assistant for both technical and non-technical users, offering a practical approach to coding challenges. Unlike traditional code generators that merely output code, OpenDevin excels by executing code directly in a console, allowing for immediate testing and verification. This functionality not only streamlines the coding process but also enhances learning and troubleshooting, making it accessible to a broader audience. In this project, we address several key challenges to improve OpenDevins effectiveness, especially in handling multi-round conversations and contextually relevant code generation. Our team identified and tackled two main challenges faced by OpenDevin: variety of input, and multi-step conversations. Through incorporating a series of functions to parse, summarize, and organize LLM agents memory logs, we significantly improved OpenDevin agents capabilities among a variety of tasks. The integration of efficient memory management led to a notable increase in accuracyfrom 44.4% to 88.9% in multi-round conversations, highlighting the importance of effective memory management in AI-powered coding tools. This report details our methodology, the challenges we faced, and the solutions we implemented, showcasing OpenDevins potential to revolutionize the way users from various backgrounds engage with coding tasks.