In a materials research lab, there are many recurring problems such as DFT convergence issues, input file errors in specific codes, and common failure patterns in postprocessing scripts, so it would be worth attaching an agent that remembers these experiences on a small scale. For example, for a specific composition family or a specific calculation code, it would be worthwhile to store “what failed before and how it was fixed” as structured memory.
At first, it is better to apply this not to all tasks, but to one or two representative workflows. It would be appropriate to check whether the accumulated memory reduces the number of tool calls in the next execution or shortens debugging time, and whether it remains reusable when moved to other materials systems.