SRAM: Spatially Registered Adaptive Memory for Persistent Video World Models

Interactive video world models must preserve previously generated places as an agent repeatedly leaves and returns to them. When history is organized around individual observations, successive visits can accumulate redundant or inconsistent versions of the same place, increasing storage while weakening earlier scene evidence. We introduce Spatially Registered Adaptive Memory (SRAM), which organizes history into persistent place records and adapts memory selection to the spatial content already stored. Camera poses associate observations from different visits with a shared record, while a correspondence-guided compression module removes local redundancy and retains additional spatial evidence without overwriting accepted memory. To learn these correspondences, we propose a Spatial Correspondence Distillation (SCD) loss that aligns local descriptors of geometrically corresponding regions across views. This design combines global spatial organization with selective reuse of the generator's existing memory, without constructing an explicit 3D representation. We further introduce PersistBench, a benchmark that evaluates how well video world models recall previously generated scenes, how stable this memory remains across repeated revisits, and how efficiently it is stored. Experiments on MBench and PersistBench show improved scene retention and lower additional peak storage relative to the compared persistent-bank baselines. Further results is available on our [anonymous project page](https://anonymous.4open.science/r/sram-persistbench-project-BEB2/README.md).