Abstract
Spatial Memory Agent
Spatial Memory Agent (SMA) studies parameter-update-free self-evolution for spatial intelligence. In a verifiable spatial environment, a frozen vision-language model receives feedback and uses verifier-guided reflection to distill compact, transferable lessons from experience. Each lesson is assigned a Transfer Reliability Score calibrated from later retrieval outcomes. During read-only deployment, semantic filtering and TRS-aware ranking retrieve reliable procedures to guide inference without changing model weights or relying on external expert spatial tools.
Experience-grounded procedural memory lets frozen vision-language agents improve spatial reasoning without parameter updates or expert tools at deployment.