Cognitive Load

Working memory is the bottleneck of learning. Miller (1956) established that humans can hold roughly seven items in working memory simultaneously. For complex, interacting elements, the practical limit is closer to three or four. Of course Tiktok wasn’t around then. Years of brainrot has reduced my brain’s holding capacity to about three things. Every design decision in a learning product either wastes this capacity, respects it, or invests it in learning.

Sweller’s Cognitive Load Theory (1988) formalizes this into three types of load that are additive. When their total exceeds available working memory capacity, learning stops.

Intrinsic load comes from the inherent complexity of what is being learned: how many elements must be held in mind simultaneously and how they interact. Translating a single word (low element interactivity) is intrinsically simpler than constructing a grammatically correct sentence in a foreign language (high element interactivity). You cannot eliminate intrinsic load without simplifying the material. You can manage it through sequencing. see the Segmenting and Pretraining pattern.

Extraneous load comes from poor design. Searching a dictionary while solving a problem, parsing a cluttered interface, reading on-screen text while hearing identical narration — none of these contribute to learning. They consume working memory that could be used for schema construction. The patterns that reduce extraneous load: Coherence (remove extraneous material), Contiguity (align related elements spatially and temporally), Redundancy (do not present the same information in competing formats), Modality (offload to the auditory channel when the visual channel is saturated).

Side-note: back in my product localization days, English speakers would always try to impose minimalizm to Chinese design, considering the apparent visual clutter on Chinese sites as a design flaw. My working hypothsis - the Chinese language has higher density and native readers are trained to hold more visual information.

Germane load is the cognitive effort directed at building and automating schemas — the structures in long-term memory that let experts recognize problem types and solution strategies without conscious deliberation. This is the load you want. Generative techniques (self-explanation, summarization, teaching others) increase germane load. But germane processing only happens when extraneous load leaves room for it.

The design implication: cognition is a budget

Clark and Mayer (2016) frame the practical consequence as three failure scenarios. Extraneous overload: bad design consumes all capacity — the learner spends working memory on irrelevant processing and has nothing left for the content. Essential overload: the material is inherently too complex for the learner’s current level, even with clean design. Insufficient generative processing: extraneous load is low and intrinsic load is manageable, but the learner passively absorbs without constructing meaning. Every design review should ask: which scenario are we in?

Reduce extraneous load first

Extraneous load has no learning value. Reduce it before doing anything else. If learners are struggling, the first diagnostic question is not “how do we make this easier?” but “what are we making them process that does not contribute to the learning objective?” Common sources in learning products: decorative images unrelated to the content, background music during instruction, split-attention layouts that force learners to mentally integrate information presented in separate locations on screen.

Side note: this doesn’t mean stripping out ALL design. MarcoPolo Learning’s World School app had landing experience where characters sat around an interactive whiteboard. If a child tapped on a character, they would interact as if tapped on the shoulder. Kids LOVED that pattern. Taste, playfulness, immersion is not waste.

Manage intrinsic load through sequencing

When material is inherently complex, the solution is not to remove complexity but to sequence it. Teach component concepts before combining them (pretraining). Break continuous units into learner-paced segments. Start with simplified whole tasks and increase complexity across task classes (Van Merriënboer and Kirschner, 2018). Skeleton before flesh. Frame before load. The goal is to build the schemas that let learners chunk multiple elements into single working memory slots, progressively freeing capacity for the next layer of complexity. See the Segmenting and Pretraining pattern and the Scaffolding pattern.

Invest freed capacity in germane processing

Once extraneous load is minimized and intrinsic load is managed, the remaining capacity should be directed toward schema construction. Prompting self-explanation, requiring learners to generate summaries, using completion tasks that demand active integration — these techniques increase germane load productively. But they only work when there is capacity to spare. Adding generative activities to an already overloaded learner produces frustration, not learning. See the Generative pattern and the Worked Example pattern.

Boundary conditions

Cognitive load is relative to the learner What overwhelms a novice is trivial for an expert, because experts have chunked elements into schemas that occupy fewer working memory slots. The same lesson design that appropriately manages load for beginners may bore or even hinder advanced learners. See the Expertise Reversal pattern.

Not all difficulty is extraneous. Bjork’s desirable difficulties — spacing, interleaving, retrieval practice — increase processing effort but improve long-term retention. The distinction: desirable difficulty targets the learning-relevant demands of the task. Extraneous difficulty targets everything else. If you cannot articulate what the learner gains from the struggle, it is probably extraneous. See the Memory pattern.

Measurement is indirect. You cannot observe cognitive load directly. Proxies include response time, error rate, self-reported effort (Mental Effort Rating scale), and physiological measures. In production learning products, learning curve analysis (tracking error rates across practice opportunities for a specific knowledge component) is the most actionable approach. A flat or rising learning curve signals that something is wrong with the load balance.

Side Note: I once observed a child frowning and acting tense throughout a lesson and thought it was the material’s problem. Turns out he was just holding back poo.