Quick Read

A mathematics lesson can contain perfectly correct mathematics and still be badly designed for learning.

The reason is simple: the learner does not receive a lesson as a complete package. The learner must hold information, connect representations, remember intermediate states, choose operations, suppress irrelevant information and update the solution while the problem is unfolding.

All of this places demands on a limited cognitive workspace.

But the newest research gives us an important correction:

The goal of mathematics teaching is not simply to minimise cognitive load.

The goal is to make sure that the learner’s available cognitive resources are being used on the mathematics that matters, rather than being consumed by avoidable searching, unnecessary representations, forgotten intermediate steps or poorly organised instruction.

This produces a better design rule:

Protect the mathematical operation. Remove avoidable interface load. Preserve useful thinking.

And there is an equally important second rule:

The same lesson does not impose the same load on every learner.

Prior knowledge matters. Familiarity matters. Representation matters. Whether information can be written down matters. The number of interacting elements matters. Even a diagram intended to make a problem easier can sometimes make it harder.

So a well-designed mathematics lesson is not merely clear.

It is receiver-aware.

The Direct Answer

Working memory matters in mathematics because mathematical problem solving frequently requires a learner to keep several pieces of information active while manipulating them.

Consider a student solving an algebraic equation.

The student may need to remember:

• the original expression;
• what transformation has just been performed;
• which terms have moved;
• whether a sign has changed;
• what equality requires;
• the next objective;
• restrictions or conditions;
• and whether the resulting line is still equivalent to the line before it.

The difficulty is therefore not simply:

Can the student perform subtraction?

It may be:

Can the student preserve the state of an entire symbolic system while performing subtraction inside it?

That distinction changes lesson design.

A learner may understand every individual operation involved yet still lose the structure of a multi-step problem.

This is why notation, worked lines, diagrams, examples and external working can sometimes function as cognitive support.

But they do not help automatically.

Recent experiments reinforce a more precise conclusion: the effectiveness of a support depends on what the learner already knows, what the task requires and how the support changes the information that must be processed.

1. Working Memory Is Not the Same Thing as Mathematical Ability

One of the easiest mistakes is to convert a cognitive mechanism into a learner identity.

A student struggles with a complex question.

Someone concludes:

“Their working memory is weak.”

That conclusion does not automatically follow.

The observed failure could arise from many different mechanisms:

• an unstable prerequisite;
• unfamiliar notation;
• incorrect route selection;
• excessive search;
• a representation-translation problem;
• incomplete conceptual knowledge;
• too many interacting elements;
• poor external working;
• anxiety or time pressure;
• an instructional interface that requires unnecessary attention switching;
• or some combination of these.

Working memory is therefore part of the learner’s operating state.

It is not a complete diagnosis.

This follows the wider MathematicsOS distinction:

Observed performance ≠ underlying capability ≠ mechanism of failure.

A wrong answer tells us something happened.

It does not, by itself, tell us why.

2. The Important Variable May Be Prior Knowledge

This is where recent research makes the article substantially more interesting.

A study by Dana Miller-Cotto and Josh Medrano examined sixth-grade mathematics learning using different instructional conditions involving problem solving, worked examples and fading—the progressive removal of worked steps.

The study was explicitly interested in whether working-memory differences changed the effectiveness of fading.

Fading produced the largest pre-to-post improvement among the conditions.

But the important result for this article is that working-memory capacity was not the significant moderator the researchers expected.

Prior knowledge was.

Learners’ starting mathematical knowledge better explained when the instructional treatment differed in effectiveness.

That gives us a major instructional correction.

It is too crude to reason:

LOW WORKING MEMORY
        ↓
MORE SCAFFOLDING

A more useful model is:

RECEIVER STATE
      ↓
Prior knowledge
Representation familiarity
Task structure
Support already available
Mathematical objective
      ↓
INTERACTING DEMANDS
      ↓
Appropriate instructional interface

The instructional question becomes:

What does this learner currently have to construct in working memory because it has not yet become available through existing knowledge?

That is a far more powerful question.

3. Why Experts and Novices Can Experience Different Problems

Suppose a teacher writes:

[
x^2-5x+6=0
]

For an experienced learner, the expression may rapidly activate a structured object:

• quadratic;
• factorisation candidate;
• two numbers with product (6);
• sum (-5);
• likely factors ((x-2)(x-3));
• roots (2) and (3);
• relationship to the graph.

Many relationships are being handled as organised structure.

For a novice, the same expression may instead appear as several separate symbols:

x
square
minus
5
x
plus
6
equals
0

The visual stimulus is the same.

The cognitive task is not.

This is why lesson difficulty cannot be read from the page alone.

It must be indexed to the receiver.

The same five lines of mathematics can represent very different processing demands depending on what has already been organised into durable knowledge.

4. Cognitive Load Is Therefore Receiver-Relative

This leads to one of the central principles of MATHCIV-012:

Load belongs to the interaction between receiver, task and interface—not to the worksheet alone.

We can represent this architecturally as:

Observed Cognitive Demand
        =
Receiver State
        ×
Task Structure
        ×
Representation
        ×
Required Coordination
        ×
Support Conditions

This is not an empirically fitted equation.

It is a design grammar.

Its purpose is to prevent a common mistake:

“This is an easy question.”

Easy for whom?

A simultaneous-equations question might be computationally straightforward for a learner whose algebra is stable.

For another learner, it requires simultaneously controlling:

• substitution;
• brackets;
• negative signs;
• expansion;
• equivalence;
• collection of like terms;
• equation solving;
• back-substitution;
• and checking.

The second learner is not merely solving the visible question.

They are maintaining an unstable prerequisite network while solving it.

5. The Lesson Itself Can Create Unnecessary Work

Some cognitive demand belongs to mathematics.

That cannot simply be removed.

Differentiation genuinely contains interacting concepts.

Trigonometric identities genuinely require symbolic control.

Functions genuinely require movement between notation, mapping and graphical structure.

But other demands are produced by the interface.

For example:

diagram on Page 1
+
formula on Page 2
+
teacher explanation
+
legend in another box
+
unfamiliar notation
+
student trying to remember the question

The learner may spend substantial processing capacity merely coordinating the sources.

That capacity is then unavailable for the mathematical relation the lesson was supposed to teach.

This is the importance of split attention.

The student must continually perform:

LOOK
↓
REMEMBER
↓
MOVE ATTENTION
↓
FIND SECOND SOURCE
↓
RECONNECT
↓
RETURN
↓
RECONSTRUCT PREVIOUS STATE

That entire loop may be educationally useless.

6. But “Add a Diagram” Is Not the Solution

This is one of the strongest 2026 upgrades to this article.

Educational design often treats visuals as inherently supportive.

Recent experimental evidence warns against that assumption.

Simon Skau and colleagues conducted a preregistered study with 81 nine-year-old children performing arithmetic tasks under text-based, visual-aid and irrelevant-visual-information conditions.

The researchers used behavioural measures together with functional near-infrared spectroscopy.

Their expectation was that useful visualisation could reduce cognitive load.

Instead, children made more errors and took longer to respond when the visual aid was present. The study also found evidence consistent with increased involvement of working-memory-related frontal regions under the visual condition.

The authors themselves place strong limits on the result.

The participants were nine-year-olds.

The mathematics involved additive situations using relatively small numbers.

The findings should not simply be transported to secondary mathematics, calculus or every kind of diagram.

But the study destroys one very useful misconception:

Visual information is not automatically cognitive offloading.

A picture can become another object the learner has to interpret.

A diagram can become another representation that must be translated.

A graph can help reveal structure—or create a second problem.

So:

MORE REPRESENTATION
≠
MORE UNDERSTANDING

The correct question is:

What mathematical relation becomes easier to perceive because this representation exists?

If there is no good answer, the representation may not belong there.

7. External Working Can Be Part of the Cognitive Architecture

Now consider the opposite move.

Instead of adding information, allow the learner to externalise information they would otherwise need to keep active internally.

For example:

Instead of mentally remembering

[
2x+3y=17
]

after manipulating another equation, write the relevant state down.

Instead of mentally carrying three substitutions, record them.

Instead of remembering where an algebraic sign changed, expose every transition.

This turns paper into part of the problem-solving interface.

Medrano and Miller-Cotto experimentally studied this idea in mathematical problem solving by comparing participants who could use paper-and-pencil offloading with participants who could not.

Both groups improved, but improvement was larger in the offloading condition by the reported effect-size comparison.

Crucially, however, the researchers did not find a simple significant interaction saying that learners with lower working-memory capacity uniformly benefited more.

Their analyses instead suggested more complicated interactions involving working-memory ranges and prior knowledge.

Again the same pattern appears:

Receiver state matters more than the slogan.

“Use working” is not a magic intervention.

But external working can reduce the need to preserve every intermediate state internally.

That is particularly important in mathematics because written notation has another advantage:

it leaves an inspectable trail.

8. Working Does Two Jobs at Once

Traditional advice often tells students:

“Show your working.”

Usually this is explained as an examination requirement.

But working performs at least two deeper functions.

Function 1: State storage

Writing preserves intermediate states outside working memory.

Instead of remembering:

What did I substitute?
What was the previous sign?
What value did I obtain?

the student can inspect the record.

Function 2: Verification

The written sequence allows the student—or another person—to inspect the transition:

[
S_0 \rightarrow S_1 \rightarrow S_2 \rightarrow S_3
]

If (S_3) is wrong, we can locate the first transition where validity was lost.

That makes written mathematics a form of external memory plus audit trail.

This is a civilisation-level property of mathematical notation.

Mathematics does not merely let the individual think.

It allows thinking to become inspectable outside the individual’s head.

That is one reason formal notation can transmit complex operations across people and across time.

This CivilisationOS interpretation is an architectural inference from the learning evidence; the cited experiments themselves did not test civilisation-level transmission.

9. The First-Invalid-Step Method

This gives teachers a much better diagnostic tool than simply marking a final answer incorrect.

Consider:

[
3(x-4)=18
]

Student writes:

[
3x-4=18
]

[
3x=22
]

[
x=\frac{22}{3}
]

There are several incorrect lines.

But there is only one first invalid transition.

It occurs here:

[
3(x-4)
\rightarrow
3x-4
]

The distributive relation was not preserved.

Everything downstream is contaminated.

So instead of correcting three errors, the teacher can locate one active failure:

VALID STATE
↓
FIRST INVALID TRANSITION
↓
DOWNSTREAM PROPAGATION

This reduces instructional noise.

And it lowers the cognitive burden of repair.

The learner does not have to hear:

“Everything here is wrong.”

The learner can see:

“Up to this point, your system was valid. This transition broke it.”

That is both mathematically more accurate and cognitively cleaner.

10. A Mathematics Lesson Is an Interface

This leads to the larger design upgrade.

We normally think of a lesson as:

Teacher
↓
Content
↓
Student

A better model is:

MATHEMATICAL OBJECT
        ↓
INSTRUCTIONAL REPRESENTATION
        ↓
RECEIVER INTERFACE
        ↓
WORKING PROCESS
        ↓
EXTERNAL WORK
        ↓
FEEDBACK
        ↓
UPDATED REPRESENTATION
        ↓
INDEPENDENT EXECUTION

The lesson therefore has an interface-design problem.

The teacher must decide:

• what to show;
• what not to show;
• what to keep visible;
• what to say;
• what the learner should reconstruct;
• what should be written externally;
• which representation should appear first;
• when a second representation becomes useful;
• when support should disappear;
• and which mathematical operation must remain inside the learner.

That last question is critical.

11. Do Not Reduce the Load by Removing the Learning

There is a dangerous version of cognitive-load optimisation.

Suppose the student must learn to factorise a quadratic.

We can reduce their immediate load dramatically by giving them the factorisation.

But now we have removed the target cognitive operation.

Likewise, an AI system could reduce load by:

• choosing the method;
• generating the algebra;
• identifying the relevant theorem;
• producing the graph;
• performing the substitution;
• and checking the answer.

The student’s cognitive load becomes very low.

So does the amount of mathematical capability they are required to produce.

Therefore:

Minimum cognitive load is not the objective.

The objective is:

Minimum unnecessary load while preserving sufficient productive mathematical processing.

In MathematicsOS language:

REMOVE:
search waste
interface confusion
avoidable memory burden
split attention
irrelevant representation
PROTECT:
concept construction
route selection
symbolic control
retrieval
translation
verification
independent execution

That is a much more defensible learning architecture.

12. The Cognitive-Load Budget

A useful lesson-planning model is to think in terms of a cognitive-load budget.

Not a numerical score.

A qualitative allocation.

Before teaching a problem, ask:

A. What must remain cognitively active?

These are the mathematical operations the learner must actually learn.

Examples:

• recognising equivalence;
• deciding which rule applies;
• constructing a representation;
• choosing a substitution;
• distinguishing gradient from intercept;
• reasoning about rate of change.

B. What can safely be externalised?

Examples:

• long intermediate calculations;
• a previously obtained expression;
• a diagram that genuinely captures the relation;
• definitions during initial acquisition;
• formulae when memorisation is not the target operation.

C. What should disappear?

Examples:

• decorative diagrams;
• repeated identical text;
• unnecessary switching between pages;
• unexplained notation;
• multiple competing representations introduced simultaneously;
• instructions that require remembering irrelevant procedural details.

This creates:

COGNITIVE CAPACITY
        ↓
Protect for target mathematics
        +
Offload safe intermediate state
        -
Remove interface waste

Again, this is an instructional architecture, not a measured universal equation.

13. A Better Sequence for Teaching New Mathematics

For a learner entering an unfamiliar mathematical structure, the following sequence is often more defensible than presenting everything at once.

Step 1 — Stabilise the prerequisite

Ask:

What existing knowledge must be available for this lesson to work?

Do not assume it from the student’s grade level.

Probe it.

Step 2 — Expose one governing relation

Show the mathematical invariant before surrounding it with many variations.

For differentiation, for example:

[
\text{change in }y
\quad\text{relative to}\quad
\text{change in }x
]

comes before a forest of rules.

Step 3 — Use the minimum sufficient representation

Do not begin with:

• verbal explanation;
• algebra;
• graph;
• table;
• animation;
• colour coding;
• real-world story;
• AI simulation;

all simultaneously.

Begin with the representation that exposes the relation most cleanly for the current receiver.

Step 4 — Keep intermediate state visible

Do not force avoidable memory tests while teaching a new mathematical relation.

Let written state carry what need not be mentally carried.

Step 5 — Model a valid route

Make the important decisions visible.

Do not narrate every trivial mental event.

The student needs to see the structure of the solution.

Step 6 — Fade support

The 2026 fading study is particularly useful here.

Support should not necessarily disappear because an arbitrary number of examples has been completed.

Its removal should respond to demonstrated starting knowledge and subsequent performance.

Step 7 — Require independent reconstruction

Eventually the representation must be rebuilt by the learner.

Otherwise we may have created:

SUPPORTED PERFORMANCE

rather than:

INDEPENDENT CAPABILITY

14. What a Teacher Should Watch

Cognitive overload is not directly visible.

So teachers need sensors.

Useful observations include:

• repeated loss of intermediate state;
• restarting a solution unnecessarily;
• overlooking conditions;
• losing signs;
• repeatedly rereading the question;
• copying without route recognition;
• switching between representations without integrating them;
• unusually high hint dependence;
• correct execution immediately after modelling but failure after a delay;
• correct method when the question type is announced but failure when mixed with other methods.

None proves “working-memory weakness”.

They are signals.

They trigger further diagnosis.

15. The Difference Between Support and Dependency

A support is successful when it helps build a capability that eventually operates with less support.

A dependency exists when the support becomes necessary for the capability to appear.

So the control loop should be:

SUPPORT
↓
SUCCESSFUL EXECUTION
↓
FADE
↓
RETEST
↓
TRANSFER
↓
DELAY
↓
INDEPENDENCE

If performance collapses after support disappears, the correct conclusion is not:

“The support worked.”

The more precise conclusion is:

“Supported performance was achieved. Independent capability is not yet established.”

That distinction protects the MathematicsOS BaseFloor.

16. This Is Especially Important in Additional Mathematics

Additional Mathematics increases the number of situations in which a learner must maintain several mathematical states simultaneously.

For example, a differentiation problem may involve:

• recognising a composite expression;
• selecting the appropriate rule;
• preserving algebraic structure;
• differentiating correctly;
• simplifying;
• interpreting the derivative;
• checking whether the answer matches the requested form.

A trigonometric equation may simultaneously require:

• identity recognition;
• algebraic transformation;
• exact values;
• quadrant reasoning;
• solution interval control;
• duplicate-solution checking.

The visible topic may therefore be “trigonometry”.

The active bottleneck may be algebraic state control.

This is why MathematicsOS treats prerequisite structure as a graph rather than a simple chapter sequence.

17. The Larger Civilisation Insight

The learning-science evidence concerns learners.

It does not empirically prove a theory of civilisation.

But it supports a useful architectural inference.

Civilisation depends heavily on the transmission of operations that no single generation should have to rediscover from zero.

Written mathematics is one of humanity’s methods for externalising cognitive structure.

A formula can preserve a relationship.

A proof can preserve a chain of justification.

A diagram can preserve spatial structure.

Notation can preserve intermediate state.

An algorithm can preserve a reproducible procedure.

A worked derivation allows another person to audit the path.

So mathematical civilisation does something powerful:

PRIVATE COGNITIVE OPERATION
        ↓
EXTERNAL REPRESENTATION
        ↓
INSPECTABLE STATE
        ↓
TRANSMISSION
        ↓
RECONSTRUCTION BY NEW RECEIVER
        ↓
INDEPENDENT USE

But the last two steps matter.

A civilisation has not regenerated a capability merely because information has been stored.

The next receiver must be able to reconstruct and use it.

That is the same distinction MathematicsOS makes between:

DELIVERY
≠
RECEIPT
≠
RETENTION
≠
TRANSFER
≠
INDEPENDENT CAPABILITY

18. The Strongest 2026 Upgrade

The older, simplified version of cognitive-load advice is easy to remember:

Working memory is limited, so make instruction simpler.

The research now supports a more sophisticated formulation.

The upgraded model is:

1. Working memory is constrained.

But observed difficulty should not automatically become a fixed working-memory diagnosis.

2. Prior knowledge changes the task.

The same mathematics can create different processing requirements for different receivers. The 2026 fading study found prior knowledge—not working-memory capacity—to be the important moderator in that instructional comparison.

3. Externalisation can help.

Writing intermediate states outside the mind can support problem solving, but the 2025 offloading experiment did not produce a simple universal low-working-memory rule.

4. Visualisation is not automatically offloading.

A 2026 experiment found that visual aids increased errors, response times and indicators associated with cognitive demand in its particular arithmetic task and population.

5. Therefore the receiver–task–representation interaction matters.

There is no universally low-load worksheet independent of the learner encountering it.

19. A Practical Mathematics Lesson Control Panel

Before a lesson:

Receiver

• What is already stable?
• What is merely familiar?
• Which prerequisites are uncertain?

Target

• What mathematical operation should become more capable today?

Representation

• Which representation exposes that operation most clearly?

Load

• What must the learner keep active simultaneously?

Offloading

• Which intermediate information can safely remain visible?

Interface

• Is attention being split unnecessarily?

During the lesson:

Sense

• Where does the learner lose state?
• Where does the first invalid step occur?
• Is the student understanding or merely following?

After the lesson:

Verify

• Can the learner reconstruct the idea?
• Can the learner solve a changed example?
• Can the learner choose the method without being told?
• Can the learner do it after a delay?
• Can the learner do it without the tutor, worked solution or AI?

Only then does immediate performance begin to become evidence of stable capability.

20. What This Article Does Not Claim

This article does not claim:

• that every learner has the same working-memory capacity;
• that working memory alone determines mathematics performance;
• that cognitive-load theory explains every learning difficulty;
• that every visual representation is harmful;
• that every diagram is helpful;
• that all instruction should be simplified;
• that difficult thinking should be removed;
• that students should never solve unfamiliar problems;
• that external working eliminates the need for mental fluency;
• or that neuroscience can identify the correct lesson design from brain activity alone.

The evidence is more useful precisely because it is narrower.

Final Answer

The best mathematics lesson is not the lesson containing the greatest amount of explanation.

It is not the lesson containing the greatest number of representations.

It is not even the lesson producing the lowest possible cognitive load.

It is the lesson in which the learner’s limited cognitive resources are directed toward the mathematical structure that must become independently controllable.

That means:

SENSE THE RECEIVER
↓
PROTECT THE PREREQUISITE FLOOR
↓
REMOVE AVOIDABLE INTERFACE LOAD
↓
EXTERNALISE SAFE INTERMEDIATE STATE
↓
PRESERVE THE TARGET MATHEMATICAL OPERATION
↓
MODEL
↓
FADE
↓
RECONSTRUCT
↓
TRANSFER
↓
VERIFY AFTER DELAY

The fundamental design principle is therefore not:

Make mathematics cognitively easy.

It is:

Make the mathematics cognitively possible without doing the mathematics for the learner.

That difference is the boundary between an explanation that merely feels clear and an instructional system capable of building mathematical independence.

Research Boundary

The recent studies discussed here concern particular populations, tasks and instructional conditions. The 2026 visualisation experiment involved 9-year-olds performing arithmetic tasks; the fading study involved sixth-grade geometry learning; and the offloading study examined mathematical problem solving under its own experimental conditions. Their results should not be converted into universal coefficients for every mathematics learner or classroom.

The MathematicsOS and CivilisationOS interpretations in this article are architectural syntheses built from those findings. They are not themselves outcomes measured by those experiments.

Research Sources Checked for This Upgrade

Miller-Cotto & Medrano, “Does working memory moderate the effect of fading on math performance?”, British Journal of Educational Psychology, 2026.

Medrano & Miller-Cotto, “Understanding working memory as a facilitator of math problem-solving: Offloading as a potential strategy”, British Journal of Educational Psychology, 2025.

Skau, Karlsson, Åsberg Johnels & Helenius, “Visualizations during arithmetic tasks hamper performance and increase cognitive load in 9-year-olds: an fNIRS study”, Frontiers in Psychology, 2026.

MATHCIV-012 · FRONTIER research article · research refreshed 10 August 2026