Does Your Agent's Memory Survive a Model Upgrade? (arXiv 2609.05339, Mon 7 Sep digest) compares four memory representations as the same history is preserved verbatim for long-context reading (LC-RAW), chunked for RAG, compressed by a model into natural-language notes (NOTES), or normalized into a fixed-schema knowledge graph (KG-fixed). Across 48 synthetic histories with randomized answer codes and exact scoring, using two open-weight models under 10B, fixed-schema structures transfer essentially perfectly across writer swaps (KG-fixed accuracy shifts by +0.0004 +/- 0.0020), while compressed NOTES are highly model-coupled — accuracy moves asymmetrically by +9.91 or -13.28 percentage points depending on migration direction. In RAG systems, a 50/50 mixed embedding index (partial migration) captures worse than either pure index.
Model upgrades are routine; memory migrations are not — and this quantifies the failure precisely: notes written by model A get reinterpreted by model B with double-digit accuracy swings, while schema-normalized memory is writer-agnostic. The engineering rule is cheap to state: normalize long-lived agent memory into an explicit schema (or raw evidence) rather than model-compressed summaries, and never run a mixed embedding index during migration — re-embed fully or pin the writer. Follows StateMemBench in making agent-memory failure modes measurable.
| Representations | LC-RAW (verbatim long-context), RAG (chunked), NOTES (model-compressed natural-language notes), KG-fixed (fixed-schema knowledge graph) |
|---|---|
| Setup | 48 synthetic histories with randomized answer codes, exact scoring, two open-weight models <10B |
| Results | KG-fixed writer swap: +0.0004 +/- 0.0020 accuracy (portable); NOTES: +9.91 / -13.28 pp asymmetric shifts (model-coupled); RAG 50/50 mixed embedding index underperforms both pure indexes |
| Rule | normalize long-lived memory to an explicit schema or raw evidence; avoid model-compressed notes and mixed embedding indexes during migration |