Controlling scientific instruments requires parameters to be adjusted progressively based on feedback from successive operations. When unobserved operating conditions evolve with the operational history, the responses become nonstationary, and a static action‑response mapping is insufficient for shot‑to‑shot decision making. Shot‑to‑shot tuning of neutral beam injection (NBI) ion sources exemplifies this challenge: electrode surface and vacuum conditions are hard to observe directly, discharge outcomes vary between shots, and experiments are costly and risky.\ We propose a physics‑grounded multimodal agent composed of three independently auditable components. Deterministic algorithms extract symbolic physical quantities and visual evidence from multichannel waveforms; a two‑layer memory organizes trajectories of the current task, experience from earlier tasks, and procedural knowledge; a multimodal large language model (MLLM) generates candidate actions based on retrieved context; and an independent feasibility check enforces action constraints before execution.\ In a strictly chronological replay of 455 shots from four EXL‑50U platforms, using physical symbols alone improved macro‑F1 by 0.21‑0.32 compared to waveform images alone. Context retrieval increased agreement with historical action directions by 22.4‑42.7 percentage points, while the contribution of visual evidence varied across models.\ Under supervised operation on five ion sources, the system achieved an 86.80% beam‑extraction rate across 409 executed shots. These results demonstrate the engineering feasibility of the framework for sequential tuning of a physical scientific instrument.\ \ Review: This work demonstrates the feasibility of multimodal agents for adaptive tuning of nonstationary scientific instruments, highlighting the advantage of symbolic information and the boost from contextual retrieval.