Organizational Memory for Agent-Guided Peptide Discovery

Scientific agents need persistent knowledge when experimental campaigns extend beyond one model context. To infuse such knowledge We construct a provenance-linked Wiki from 34 papers and freeze it before evaluation. We test whether read-only access to this literature memory improves finite-budget discovery over 409 measured sequence--pH candidates for peptide amphiphiles. The Wiki and no-Wiki campaigns have mean cumulative-hits scores of $15.24\pm3.83$ and $17.04\pm3.88$ hits, respectively, across ten campaigns per condition. The difference is $-1.80$ hits ($95\%$ bootstrap CI $[-4.93,1.51]$, exact permutation $p=.304$), providing no evidence that Wiki access improves discovery. An exploratory trace audit suggests a task-alignment failure mode. All campaigns select the same unsuccessful candidate within $20$ measurements, and Wiki campaigns selected it earlier, at median measurement $3$ compared with $6$ without the Wiki. The campaign reports produced $30$ candidate memory-update plans, of which one chemist rated $8$ as useful and $3$ as partially useful. These results identify task alignment and campaign-level calibration as requirements for scientific memory, while showing that campaign experience can yield useful proposals for a curated literature knowledge base.