FocusMem: Factorizing Content, Readout, and Trust in Latent GUI Memory
FocusMem tries to make GUI-agent memory selective, not just smaller.
The paper argues that fixed latent memory blocks can lose key details, serve the wrong decision stage, or pull in irrelevant past trajectories. Its proposed system separates what memory stores, how the current state reads it, and whether the agent should trust it at a given step. The GUI policy stays frozen while those memory components are trained. Across five GUI-agent benchmarks, the authors report gains over matched fixed-memory and prior latent-memory baselines. HF Daily Papers' note
The paper argues that fixed latent memory blocks can lose key details, serve the wrong decision stage, or pull in irrelevant past trajectories. Its proposed system separates what memory stores, how the current state reads it, and whether the agent should trust it at a given step. The GUI policy stays frozen while those memory components are trained. Across five GUI-agent benchmarks, the authors report gains over matched fixed-memory and prior latent-memory baselines. HF Daily Papers' note
score 4