Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence
A frozen VLM can improve spatial reasoning by reusing verified experience as runtime memory, without retraining or external spatial tools.
Spatial Memory Agent stores compact “lessons” distilled from verifier-guided reflections in a spatial environment.
Each lesson gets a Transfer Reliability Score, which is updated based on later retrieval outcomes.
At deployment, the system retrieves relevant lessons to guide inference while leaving the base model unchanged.
Across five spatial benchmarks and four VLMs, the paper reports the best macro average in every base-model block.
HF Daily Papers' note
Spatial Memory Agent stores compact “lessons” distilled from verifier-guided reflections in a spatial environment.
Each lesson gets a Transfer Reliability Score, which is updated based on later retrieval outcomes.
At deployment, the system retrieves relevant lessons to guide inference while leaving the base model unchanged.
Across five spatial benchmarks and four VLMs, the paper reports the best macro average in every base-model block.
HF Daily Papers' note
score 5