EvoMem: Persistent Memory for LLM-Based Evolutionary Code Optimization

A recent preprint on arXiv presents EvoMem, an innovative persistent memory architecture aimed at improving LLM-driven evolutionary program searches by capturing and reusing knowledge from candidate mutations across various tasks and runs. The document, referenced as arXiv:2608.10795v1, posits that effective mutation strategies in evolutionary code searches often harbor reusable insights that transcend individual runs and can be applied to related domains. Current LLM-driven evolutionary models tend to overlook this valuable knowledge, resulting in the redundant rediscovery of similar concepts and restricting cross-run and cross-task learning. EvoMem remedies this by transforming successful mutation occurrences into organized, context-aware guidance for future iterations. The architecture functions in two stages: it gathers and archives promising ideas post-run and retrieves pertinent instructions for guiding mutations in subsequent evolutions. Evaluations were conducted in areas such as geometric optimization, multi-hop question answering, and GPU kernel optimization, though specific findings are not detailed in the abstract. This research is significant for artificial intelligence, machine learning, and automated code optimization, having been published on arXiv on August 26, 2025.