Xiaohua Jia

Large Language Model (LLM)-based multi-agent systems have shown promise for automated vulnerability detection, but our industrial deployment reveals two limitations: 1) each agent's prompt is typically optimized in isolation, degrading global end-to-end performance; and 2) evolutionary prompt search lacks memory across iterations, wasting computation costs. To address these limitations, we propose MemCo, a memory-guided prompt co-evolution framework applied to industrial vulnerability detection. First, MemCo takes an end-to-end prompt co-evolution strategy, which evaluates prompts based on the system's final output rather than assessing each agent in isolation. In each evolutionary iteration, only the prompts of agents responsible for the misclassification undergo further evolution, thereby enhancing the system's global accuracy. Second, MemCo adopts a memory-guided search mechanism, which edits prompts based on records of successful and failed modifications. This reduces redundant blind exploration and significantly lowers computation costs. Comprehensive experiments show that MemCo improves Macro-F1 by +8.4 to +18.0 points on public benchmarks and up to 36.9% relative improvement on a proprietary industrial benchmark, while reducing per-generation evolution tokens by 48.5% and inference tokens by 63.8%. MemCo is deployed as a vulnerability detection assistant in our production security pipeline.