Definition
Plain language
How much information someone has to juggle in their head at once.
As stated in the literature
In multi-step reasoning, the working-memory demand placed by a task, often used as analogy for context-window or hidden-state utilization in models.
Why it matters: The concept maps neatly onto how much state a model can usefully track in its context, which shapes how tasks should be structured for agents.
For example, a recipe that interleaves ten ingredients and three timers imposes higher cognitive load than one that handles them in clean sequential blocks.