Memory has become central to self-evolving LLM agents, which store and retrieve experience from past rollouts to guide future decisions. Existing systems represent this experience as free-form natural language and maintain it by rewriting, which fits tasks whose experience is narratable.\n\nIn scientific discovery and biomolecule optimization, experience is often a quantitative measurement provided by an oracle under a specific context. Rewriting it as text loses the index for a record, replaces a measurement with a description, and cannot express how interventions interact.\n\nTo address this, we introduce the MEMIG framework, arguing that memory for biomolecule optimization should be indexable, quantitative, and structured. MEMIG uses domain‑grounded identifiers to give each record a unique index and accumulates the quantitative effect measured by a property oracle, keeping updates local and retrieval an exact lookup. It then connects interventions applied together and measures their epistasis by decomposing each evaluated set.\n\nResults on various in‑silico antibody optimization benchmarks show that MEMIG improves performance in most settings.\n\nReview: MEMIG demonstrates notable performance gains in antibody optimization, highlighting the importance of structured memory for scientific discovery.