Multi‑agent systems built on large language models (LLMs) tackle complex reasoning by configuring several agents and orchestrating their collaboration. The main challenge is to adapt the orchestration as the collaboration state evolves. Routing decisions based only on the query lack awareness of intermediate progress or errors, hurting accuracy. Using the full execution history as context supplies the missing information but forces every subsequent decision to process all prior steps, including redundant or low‑utility ones, leading to an execution‑history overload and higher cost. Effective orchestration therefore requires a compact state that captures useful progress without accumulating unnecessary context. We introduce Gated‑Memory Routing, which conditions each decision on both the query and a learned execution memory. A Memory Write Gate commits only non‑redundant reasoning steps to memory, while a Retrieval Gate provides each agent with a concise, relevant subset, ensuring decisions are based on a clean, informative state. At each step the system selects the next role and backbone from this memory, and an Adaptive Halting Controller stops execution once the memory contains sufficient evidence to answer. Across five reasoning and code‑generation benchmarks, our framework achieves the best average accuracy, surpassing the strongest baseline by 2.44 points, and reduces HumanEval inference cost by 31.9% relative to that baseline. Code is available at https://github.com/rajibrhasan/gated-memory-routing
Review