What Should an ADMET Design Agent Remember? A Learned Experiential-Memory Tool for the Analyze Step of the DMTA Loop

Design agents that drive the design–make–test–analyze (DMTA) loop of lead optimization need a memory of past optimization decisions to evaluate the next structural edit; today they retrieve isolated matched molecular pairs, or nothing. We present CETR (Context-Engineered Transformation Reasoning), a learned experiential-memory module for such agents: it stores multi-step lead-optimization trajectories and reads a budgeted prior-in-series context into a grounded evaluation of a candidate transformation's multi-endpoint ADMET consequences—the analyze step an LLM controller would invoke as a tool. We formalize the read as a single differentiable operator, $\mathrm{MemRead}_\theta = \mathrm{Compact}\circ\mathrm{Clear}_B\circ\mathrm{Retrieve}$, whose ablations (empty store, uniform compaction, no learned weighting) are limiting cases of one operator rather than unrelated baselines. On 27,588 trajectories spanning 21 ADMET endpoints (116,999 transformations from 54,205 ChEMBL 37 compounds), the full read reaches macro Pearson $r=0.832$ against $0.554$ for a retrained multi-task D-MPNN—a large $+0.28$ macro gain we attribute honestly: $+0.27$ is the framework's $\Delta$-parameterization (predicting changes directly rather than subtracting two absolute predictions, so an empty memory already reaches $0.827$), and only $+0.005$ is reading a populated memory on this fingerprint backbone ($+0.04$ on a learned D-MPNN encoder). We treat this last increment as small—the learned budget weighting is statistically indistinguishable from uniform ($0.832$ vs. $0.833$)—bounding what experiential memory is presently worth rather than overclaiming it. On this corpus the context-management primitives—budgeted clearing and query-conditioned compaction—shift macro $r$ by at most $\sim$0.02; we show why (retrieved context rarely exceeds the budget for short within-series trajectories) and when they should re-engage. We do not ship the controller, but show precisely where this memory module plugs into the DMTA loop.