Fixed-schema agent memory survives a model swap almost exactly; model-written notes swing 23 points depending on migration direction

arXiv 2609.05339·high signalXBlueskyLinkedInCopy link

A controlled study across 48 synthetic histories and two sub-10B open-weight models compared four memory formats through a writer swap. A fixed-schema knowledge graph moved by +0.0004 ± 0.0020 accuracy, while compressed natural-language NOTES shifted +9.91 or -13.28 points depending on which direction the migration ran, and a 50/50 mixed embedding index captured only 4.96 of the 11.90-point gain from full re-embedding. Store-only repair of NOTES failed to reach 90% recovery in all 48 cases, while keeping the raw source history recovered 34 of 48, so the practical rule is: normalize memory to a schema, never partially re-embed, and keep the raw transcripts you compressed from.