MaxMind: A Memory Loop Network to Enhance Software Productivity Based on LLMs
Yuchen Dong, Xiaoxiang Fang, Yu‐Chen Hu, Renshuang Jiang, Zhe Jiang · 2024
Large language models can help facilitate automated software operations and tool generation (SOTG), thus augmenting software productivity. Current research often overlooks the significance of converting real-time task experiences into system memory and fails to recognize the pivotal role of differentiating the knowledge value for future reference. This paper provides a novel system model, MaxMind, to address these issues through two novel designs: (1) evolving external memory models into Memory-Loop Networks for timely memorization and experience referencing, and (2) enhancing a RAG mechanism with knowledge precision segmentation to utilize memory based on value differentiation. To demonstrate our approach, we developed MaxMind4Sheet, an electronic spreadsheet processing system that aligns with the MaxMind philosophy. Comparative experiments with SheetCopilot have demonstrated that the accumulation and recycling of task memories lead to a steady enhancement in the task success rate, with an improvement rate of approximately 3%-6% per round in this implementation example. Note that as the memories continue to accumulate, this cumulative improvement may become substantial. These suggest that MaxMind has significant potential to enhance the capabilities and productivity of LLM systems in SOTG.