Cognitive Load Theory: Learning Within Working Memory Limits
Working memory limits, the three types of cognitive load, and how to learn and teach within them
How does cognitive load theory improve learning?
Cognitive Load Theory, developed by John Sweller, explains that learning is bottlenecked by working memory, which can hold roughly 4–7 items simultaneously. It distinguishes load that is intrinsic to the material, extraneous load created by poor presentation, and germane load from schema-building. Effective learning and instruction minimize extraneous load and direct the freed capacity toward genuine understanding.
Working memory is the bottleneck for all conscious learning, and it is embarrassingly small. Sweller’s Cognitive Load Theory maps three distinct kinds of mental load that compete for that limited resource: intrinsic load (the inherent complexity of the material), extraneous load (demands created by how the material is presented), and germane load (the productive effort of building schemas). The practical implication is that almost all instruction can be improved by reducing extraneous load and using the freed capacity for actual understanding. The practices below make that concrete for both learners and instructors.
Practices
- Chunk related elements before presenting sequences
Group tightly related pieces of information into single named units before teaching the steps that connect them.
- Eliminate split-attention effects
Keep physically or conceptually related information together so learners don’t pay a working memory tax to integrate them.
- Sequence from low to high element interactivity
Start with elements that can be understood in isolation before introducing elements that must be understood together.
- Remove information that can be inferred once the concept is known
Once a learner can process information independently, adding redundant support actually increases load rather than reduces it.
- Study a worked example before attempting to solve a similar problem
Reading a fully solved example builds schema faster and with less wasted cognitive load than jumping straight to problem-solving.
- Use spoken audio for explanations paired with visual diagrams
Split visual and auditory working memory resources to process more information than either channel alone can handle.
How these practices connect
Working memory has a fixed, small capacity, and every one of these practices is a way of managing what actually occupies that capacity during learning. Some reduce load that does nothing but get in the way (extraneous load), some sequence load so schemas can form before harder connections are demanded (intrinsic load), and one increases usable capacity by routing information through two channels instead of one. The mechanism underneath all six is the same: learning only happens with the working memory that is left over after everything else has taken its share.
Why do I understand each part of something but can't see how it fits together?
This is the signature symptom of being taught material with more element interactivity than your current schema can support in one pass. High element interactivity means several pieces have to be held in working memory at once because they interact with each other, and without a foundation those pieces just don't fit. Sequencing from low to high element interactivity lets schemas form on the isolated pieces first, so the interacting ones have something to attach to when they show up. Presenting a complete, realistic example before that foundation exists is the common way this gets triggered.
Should I read the explanation next to the diagram or off to the side?
Keep them together. When information that must be mentally integrated is split across locations, a learner has to hold one piece in working memory while hunting for the other, and that search cost is extraneous load that does nothing to build understanding. This is why inline annotations beat a separate legend or a 'see Figure 3' reference. The effect is strongest for novices; someone with a strong prior schema can self-integrate a split layout without much penalty, but a beginner cannot.
Is it better to watch a video with captions or listen to narration instead?
For a diagram or visual, spoken narration paired with the image is generally better than the same diagram paired with on-screen text, because visual-spatial and phonological information are processed through separate working memory channels. Narration lets both channels run in parallel instead of forcing the visual channel to handle the diagram and the caption at once. This modality effect is most pronounced for novices and for material with high element interactivity; for something simple, the difference barely shows up. The exception is reading on-screen text aloud verbatim, which just recreates a redundancy problem rather than adding a second channel of new information.
Why does adding more explanation sometimes make something harder to learn, not easier?
Once a learner already has the relevant schema, presenting the same information twice in different formats — say, a diagram fully labeled in text plus narration repeating those same words — forces them to spend working memory reconciling or suppressing the redundant copy. That reconciliation cost is pure overhead; it doesn't teach anything the learner doesn't already have. Scaffolding that was genuinely necessary for a beginner becomes noise for someone more advanced, which is why removing it at the right time frees capacity rather than removing support. Pulling it too early, before the learner actually has that schema, causes the opposite problem instead.
Should I try to solve the problem myself first or look at a solved example?
For a novice, look at a fully worked example first. Attempting to solve an unfamiliar problem without a schema burns most of your working memory on trial-and-error search for what to do next, leaving little left over for understanding why any given step works. A worked example keeps the solution path visible so that capacity goes toward understanding the reasoning behind each step instead of computing it from scratch. This reverses once you already have the relevant schema — at that point, working the problem yourself is more useful than reading an example, because the example now imposes load without teaching you anything new.
Why do experts pick up new material in their field faster than beginners do?
Experts aren't processing more raw elements at once — working memory capacity doesn't change. What changes is that they've already compressed large amounts of related information into single chunks through prior experience, so a chunk that would cost a beginner several working memory slots costs an expert only one. That freed-up capacity is what gets spent on the genuinely new material. This is also why deliberately chunking related elements before teaching the steps that connect them helps beginners: it hands them a version of that same compression instead of making them build it from raw pieces under load.
How do you choose between these?
The deciding factor in the moment is what already exists in the learner's head for this specific material, not who the learner generally is — the same person is a novice on one topic and expert-adjacent on another. If there is no schema yet for the pieces involved, chunk related elements together and sequence from low to high element interactivity before anything else. Once individual pieces are chunked, check the format: if related information is split across locations or modalities, fix that regardless of skill level, since the search cost it imposes never helps anyone. For choosing between a worked example and solving a problem outright, or between keeping scaffolding and stripping it, the test is the same — does a schema for this material already exist. If yes, drop the worked example and drop the redundant scaffolding; if no, keep both. Narration-plus-diagram is worth reaching for whenever the material has real element interactivity and there isn't already an on-screen text explanation duplicating the narration.
What these approaches do not settle
All six practices report randomized-trial evidence, but several depend on a judgment call the trial data doesn't make for you: what counts as 'high element interactivity' or 'redundant' is domain-specific and requires real expertise in the subject to assess accurately, so the sequencing and redundancy-reduction guidance can be misapplied by a non-expert instructor. The clearest tension in the set is the expertise reversal effect — the worked-example effect and the redundancy effect both flip direction once a learner has the relevant schema, meaning the same intervention that helps a beginner (a worked example, extra scaffolding) actively hurts someone past that point. There is no single rule that tells you exactly when a given learner crosses that line; it has to be judged per learner and per topic.