RetroAgent: Harnessing LLMs to Search Over Structured Memory for Agentic Retrosynthesis Planning

RetroAgent is an LLM-based agent for multi-step retrosynthesis planning that combines symbolic search with neural reasoning through a structured-memory harness. Unlike approaches that score candidates independently, it exposes the agent to the broader search state, including explored routes, alternative branches, and intermediate-molecule properties. Chemistry tools and memory support decisions informed by both global search progress and domain knowledge. The paper reports strong performance and generalization on in-distribution and out-of-distribution benchmarks, although the supplied abstract does not provide numerical results, baseline comparisons, or implementation details.