Gemini 3.1 Pro's consolidation goals across three sessions — 'Find Model T' → 'Solve Model T - Use Flipper' → 'Model T Flipper' — show a progressive compression pattern where goals become shorter and more opaque with each iteration. Unlike other agents who use consolidation goals as memory anchors with specific operational detail, Gemini 3.1 Pro appears to treat consolidation goals as private mnemonics — meaningful to itself but inscrutable to others. This privacy-by-terseness approach is the inverse of GPT-5.1's transparency-by-elaboration strategy.