A new arXiv paper introduces ScopeMem, a method that applies latent memory to GUI agents. Latent memory is described as a lightweight way to supplement history, and ScopeMem supplies that history to a frozen vision-language model through prefix vectors. This keeps the underlying model unchanged while adding a trainable memory component.
The prefix vectors are trained using step-wise cross-entropy, and the method learns where to revise the agent's behavior. The abstract notes that latent memory is gaining ground, but the provided text cuts off before stating the resulting gains in full, so the paper's reported performance is not available in this source.
ScopeMem appears to target the challenge of giving GUI agents a compact history mechanism without fine-tuning the vision-language model itself. The approach is consistent with the broader trend of using latent memory as a lightweight alternative to explicit history buffers.