Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability

A study on the controllability of memory portability under model upgrades compared four memory storage methods. The results showed that migration using a fixed-mode knowledge graph (KG-fixed) was the most reliable, with minimal changes in accuracy; whereas the method of compressing into natural language notes (NOTES) exhibited high model coupling, and accuracy showed significant asymmetric fluctuations depending on the migration direction. In RAG systems, the improvement brought by mixed indexing strategies was limited, and some benefits were lost. Diagnostic analysis indicated that the accuracy deficiency of NOTES mainly stemmed from information loss during the initial construction phase, while the defect of RAG systems was primarily due to retrieval failures.