This study compares four common formats when the underlying language model is upgraded. The formats tested are raw history (LC-RAW), retrieval-augmented generation chunks (RAG), model-written notes (NOTES), and fixed-schema (KG-fixed). The authors used 48 synthetic histories with verifiable answer codes and two open-weight models under 10 billion parameters.
Fixed-schema knowledge graphs transferred reliably with virtually no accuracy change. Compressed notes lost up to 13 percentage points depending on migration direction. Partial embedding migrations in RAG systems recovered less than half the gain of full re-embedding. Store-only repair failed to recover performance in all cases, while retaining raw history enabled recovery in most test cases.
The findings are interesting because model upgrades are routine in production systems, yet memory portability is not commonly tested. The results suggest that memory format choices have significant practical consequences for reliability during model transitions.