What's this about?
Spatial intelligence is foundational for embodied agents, robotic planning and multimodal assistants. Existing approaches split into post-training the vision-language model, or hand-writing spatial reasoning pipelines. This paper (40 upvotes) proposes a third route: accumulating procedural memory grounded in experience.
What exactly is "spatial intelligence"?
It's the ability to reason about where things are and how space works — that the cup is behind the laptop, that you can't reach the far shelf from here, what the room looks like from the other side. Humans do this so automatically it's hard to see as a skill, but vision-language models are notably bad at it: they can describe a scene fluently while getting the relationships between objects wrong. The gap is that describing an image and reasoning about the space it depicts are different problems.
Why is this worth noticing?
- Procedure memory is a distinct idea from fact memory — it stores *how to go about* a spatial task learned from experience, rather than storing facts about a particular room.
- It sidesteps the two expensive options: no retraining run, and no hand-built pipeline that only covers what its author anticipated.
- It's part of a clear cluster this week — MindMemOS, Databricks' ontologies, this — all arguing that what agents lack is durable, structured, accumulated context.