MemGrad: A Memory-Guided Optimization of Agentic Software Development via Abstracted Textual Gradients

Agentic systems built on large language models increasingly operate in settings that demand stable reasoning, effective collaboration, and reliable adaptation. Existing optimization methods offer valuable signals through prompting strategies, alignment techniques, decentralized coordination, and experiential retrieval, but they do not translate feedback gathered across multiple trajectories into persistent improvements in agent behavior. We introduce MemGrad, a memory-guided optimization framework that uses textual gradients to transform batches of behavioral feedback into coherent and interpretable improvement directions. These gradients support a retrospective–prospective memory structure: retrospective memory captures recurring patterns and common failure modes, while prospective memory encodes gradient-derived strategies that guide future reasoning and coordination. The framework also updates system prompts so that agents internalize these improvements without model fine-tuning. Applied to AgileCoder, a multi-agent software development framework, our approach improves task success, reasoning stability, and alignment with user intent. These results show that text-based, memory-centered optimization provides a practical and scalable route toward more reliable agentic systems.