Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability
As base models update, an agent's stored memory does not automatically carry over. Across 48 controlled synthetic histories on sub-10B models, this study tracks how different memory formats survive a model swap. While fixed-schema knowledge graphs transfer with virtually zero drift (+0.0004 accuracy shift), natural language notes exhibit asymmetric swings of up to 13.28 percentage points, largely due to initial compression loss. Partial embedding migrations similarly forfeit most retrieval gains, and memory repair consistently fails without the raw source history.
Why it's worth reading
As LLM upgrades become routine engineering operations, this paper provides clear empirical evidence against blindly reusing natural-language agent notes, clarifying why retaining raw interaction logs and schema-bound representations is vital.