Vision-language-action models hit a memory wall, and three labs are trying to break it

arXiv hosts the most cited of the three new papers. The proposed architecture splits memory into a fast working buffer (high resolution, short horizon) and a slow episodic store (compressed, long horizon). Reported results show roughly 3x improvement on long-horizon benchmarks.

The second approach, from a different group, externalises memory entirely into a queryable database. The model writes to it and reads from it, the way a human might write notes during a task. Less elegant, but it generalises across model sizes.

The third approach changes the training procedure rather than the architecture. By training on tasks that explicitly span longer time scales, the model learns to maintain task state inside its existing attention window. Less computationally expensive but harder to scale.

What is likely next

Expect at least one of these approaches to land inside a production VLA release within six months. Expect "long-horizon" to become a marketing word the way "general purpose" is today. And expect the next set of benchmarks to be much harder, because the current ones are not capturing what real robots need.

The wall is real. The runs at it are encouraging. The field is moving.

Isaac Mendez · 23 May · 3 min