Domain experts understand business constraints but often lack operations‑research expertise or dedicated support. LLM‑based optimization agents can translate natural‑language requirements into mathematical models or solver programs, allowing existing optimization tools to run directly. This lowers the barrier to optimization, yet real‑world operations are dynamic: demand, resources, and priorities change, requiring rapid updates to data, constraints, and objectives. Methods that handle isolated requests struggle to preserve earlier decisions and reuse previous search results.
We introduce MAPLE (Memory‑Augmented Planning with Language and Evolution), an agent that maintains optimization problems through successive natural‑language requests. MAPLE combines language‑driven problem construction, mathematical programming, and evolutionary search, while retaining the full optimization program, accepted plans, prior updates, and candidate solutions for reuse in later requests.
To evaluate MAPLE we built the NLDO benchmark, consisting of 15 trajectories and 180 updates covering selection, scheduling, rostering, routing, and cloud‑resource placement. MAPLE completed all trajectories, achieving an online scalar quality of 0.951 and a Pareto hypervolume ratio of 0.875. Controlled comparisons show that preserving executable state improves update validity and can retain useful search information even after substantial revisions.
Review: MAPLE’s memory mechanism enables cross‑request knowledge accumulation, markedly improving adaptation speed and solution quality in dynamic optimization settings, and offers a practical path toward self‑service optimization for business users.