Fixed-schema memory survives a model swap; model-written notes swing 23 points.

Across 48 synthetic histories and two sub-10B open-weight models, a fixed-schema knowledge graph moved by +0.0004 ± 0.0020 accuracy through a writer swap, while compressed natural-language notes shifted +9.91 or -13.28 points depending on migration direction (arXiv 2609.05339). A 50/50 mixed embedding index captured only 4.96 of the 11.90-point gain available from full re-embedding, so partial re-embedding is close to worthless. Store-only repair of notes never reached 90% recovery in any of the 48 cases; keeping the raw source history recovered 34 of 48. Three rules fall out: normalize to a schema, never partially re-embed, keep the transcripts you compressed from.