Memory portability isn't automatic: compressed formats couple tightly to specific models and fail asymmetrically during upgrades, while structured knowledge graphs and raw histories are more robust. Always test migrations in both directions and keep original source data for recovery.
When you upgrade an AI agent's model, its memory often breaks in unexpected ways. This paper tests four memory storage formats—raw text, chunked retrieval, compressed notes, and structured knowledge graphs—to see which survive model swaps. Fixed schemas stay reliable, but compressed notes and retrieval systems lose accuracy unpredictably depending on which direction you upgrade.