LLM agents optimize a ligand in a protein pocket by editing it step‑by‑step and docking each intermediate. Recent agents first perform extra dockings to probe the complex with controlled edits; the guidance extracted from these probes is reported to be more useful than the chemical knowledge already embedded in the model. The probes are then discarded together with the instance that generated them, so a natural next step is to store them in an experience base, allowing a new pocket to inherit the work already paid for.
An experience base is accessed through a key: past edits are pooled under recorded properties and a new edit looks up the same key. Using the full log of one run over the CrossDocked2020 test‑set pockets, we ask whether its keys recur on a pocket never seen before and still separate improving from worsening edits. When the contacts at the edited site, the action, and the iteration round are used as a key, the base separates the two groups sharply.
Our analysis shows that this separation is largely explained by how much the edit changed the molecule’s size, the action itself, and the round—none of which directly describe the target pocket. This does not mean the target is irrelevant: using the pocket name outright yields an even sharper separation. For a pocket absent from the base, transfer must rely on a description rather than its identity. Among the exact‑match descriptions we evaluated, the more specific ones rarely recur, while the coarser ones provide little separation; the signals that do transfer mainly describe the edit and the resulting molecule rather than the target.
Under the representations and lookup scheme studied here, an experience base appears more useful for prioritizing an agent’s own proposals for docking than for directly prescribing what to build.
Review