Memory framework doubles GPT-5.2 success on 49 materials tasks without updates

The framework records observations, failure boundaries, protocols and validation checks as retrievable, revisable entries. It was tested in equation-of-state calculations and 13 simulation workflows covering band gaps, phonons, vacancies and work functions. Execution traces show that remembered guardrails converted initialization failures into pre-execution checks and halved token counts by the third iteration. Data from the 49-question set records baseline GPT-5.2 performance improved to 25/2/0 Correct/Partial/Error from 22/1/4 while avoiding 92 percent of repeated errors. Aggregate tool calls fell by more than half once skills migrated across runs. Outputs remained physically consistent in all preserved cases. Integration into daily materials workflows exposes a gap the paper understates: memory must interface with existing job schedulers and electronic lab notebooks rather than replace them. Portability across model stacks reduces vendor lock-in but requires standardized schema for failure facts that current agent libraries do not enforce. Operational deployment will hinge on audit logs that let human researchers approve or edit stored skills before they propagate. Labs running repeated DFT or phonon calculations can expect measurable reductions in retry overhead once memory entries are versioned and shared.