AI Isn't Outthinking Mathematicians. It's Out-Remembering Them

Davide Piffer

Aug 04, 2026

At the 1952 dedication of the Institute for Advanced Study computer. AI may be less like an electronic Einstein than a machine-amplified von Neumann: immense speed, breadth and symbolic memory.

When an AI system solves a difficult mathematical problem, the usual explanation is that it has become more intelligent.

Perhaps it has absorbed millions of mathematical examples. Perhaps reinforcement learning has taught it better reasoning strategies. Perhaps it is beginning to develop something resembling genuine mathematical intuition.

All of these explanations may contain some truth. But they overlook a simpler possibility:

AI has access to a vastly larger working memory than the human brain.

Or, more precisely, it has access to an enormous external symbolic workspace that performs many of the functions that working memory performs in humans.

This difference may be especially important in mathematics.

A human mathematician can hold only a small number of unfamiliar elements in mind simultaneously. An AI model can keep the entire problem statement, hundreds of intermediate equations, several abandoned approaches, definitions, constraints and earlier conclusions inside its context window.

We normally interpret the resulting performance as evidence of superior reasoning. But some of it may instead reflect the removal of one of the most important biological limits on human reasoning: our extremely restricted working-memory capacity.

Mathematics is constrained by memory

Working memory is the mental system that allows us to hold and manipulate information over short periods.

When solving an equation, you must remember what each variable represents, which operations have already been performed and what the current goal is. During a proof, you may need to keep track of assumptions, intermediate lemmas, exceptions and multiple possible cases.

Human working memory is remarkably limited.

Its exact capacity depends on the task and on how information is organized, but the general limitation is obvious from everyday experience. Try multiplying two three-digit numbers in your head. The underlying operations are simple. The difficulty comes largely from having to preserve partial results while performing additional calculations.

Writing the numbers down transforms the problem.

Paper does not make you more intelligent. It expands your effective working memory.

The same principle applies at higher levels of mathematics. A mathematician uses notation, scratch paper, diagrams and previously written lemmas not merely to communicate the solution, but to make the reasoning cognitively possible.

Experts compensate through “chunking.” A novice sees a long sequence of symbols. An expert recognizes a familiar structure and treats it as a single conceptual object. This allows far more information to fit inside the same biological working-memory limit.

But chunking does not eliminate the limit. It merely compresses the information.

An AI model faces a very different constraint.

Working memory predicts mathematical performance beyond IQ

The importance of working memory for mathematics is not merely theoretical. It is visible in the differences between human beings.

Working memory is strongly related to general intelligence, which raises an obvious question: does it independently predict mathematical performance, or is it merely another imperfect measure of IQ?

Several studies suggest that it contributes something beyond conventional intelligence measures. Alloway and Passolunghi (2011), for example, examined working memory, verbal ability and mathematical skills in children. They found that working-memory measures made a distinct contribution to mathematical performance rather than simply reproducing the association between mathematics and general verbal ability.

In a separate six-year longitudinal study, Alloway and Alloway (2010) measured children at age five and then examined their academic achievement six years later. Early working-memory performance predicted later literacy and numeracy even after IQ was included in the analysis. Indeed, working memory was a stronger predictor of the later academic outcomes than the IQ measure used in the study.

Blankenship and colleagues (2015) similarly reported that working memory explained unique variation in mathematical fluency and calculation after statistically controlling for IQ and age. A large meta-analysis by Friso-van den Bos and colleagues (2013) also found a consistent relationship between working memory and mathematics across primary-school studies, although the strength of the relationship varied according to the type of working-memory and mathematical task being measured.

These findings should not be exaggerated. Working memory and intelligence overlap substantially, and statistical control cannot perfectly isolate them as independent psychological mechanisms. Nor does the evidence imply that commercially training working memory will necessarily produce large improvements in intelligence or mathematics.

The narrower conclusion is nevertheless important: among children with similar measured intelligence, differences in the ability to hold, update and manipulate information still predict differences in mathematical performance.

This provides a crucial clue for understanding AI. If human mathematical performance is partly capped by a working-memory bottleneck, then giving a machine an enormous symbolic workspace changes the nature of the contest. The machine may appear more mathematically intelligent partly because it is much less constrained by a cognitive limitation that suppresses human performance.

The context window is a gigantic notebook

A modern language model can process an enormous sequence of tokens at once. This sequence may include the original question, definitions, examples, intermediate calculations and the model’s own earlier reasoning.

The context window is not identical to human working memory. It is better understood as a gigantic external notebook combined with an imperfect system for searching and using what has been written in it.

This distinction matters.

Humans possess a form of active internal memory. We can silently choose a number, hold it in mind, transform it and replace it with a new value without saying or writing anything.

Standard language models are much weaker at maintaining this kind of private, continuously updated mental state. Their most stable form of memory is usually the sequence of tokens that has already been generated.

If the model writes:

x=6

and later writes:

x+3=9,

those statements remain inside the context. The model can attend to them again when generating the next step.

Its reasoning is therefore often externalized. The text is not merely a report of a completed thought process. The text is part of the mechanism by which the reasoning occurs.

Humans do something similar when using scratch paper. The major difference is scale.

An unaided human may struggle to keep five unfamiliar conditions active simultaneously. An AI can preserve dozens or hundreds of them in explicit form.

This does not mean that every item in a long context is retrieved perfectly. Models can overlook relevant information, become distracted or lose track of details. Advertised context length is not the same as perfectly usable memory.

Nevertheless, the difference in potential capacity is enormous.

Why this matters particularly for mathematics

The context-window advantage is not equally useful in every kind of reasoning.

It matters especially for mathematics because mathematical reasoning can be translated unusually well into explicit symbols.

Almost every relevant element of a mathematical problem can be written down:

the assumptions;

the definitions;

the known equations;

the current objective;

the results already proved;

the cases that have been eliminated;

the conditions under which each step remains valid.

Once written, this information remains stable.

If (x) is defined as an integer at the beginning of a proof, it remains an integer unless the proof explicitly changes the definition. A strict inequality does not gradually become a non-strict inequality because of changes in mood, context or interpretation.

Mathematical symbols are designed to reduce ambiguity.

This makes mathematics almost perfectly suited to an intelligence that operates through a large textual workspace.

Consider a problem requiring the solver to remember that:

n is odd;

p is prime;

x≠0

and that one branch of the argument has already produced a contradiction.

A human may understand the basic strategy but divide by x before establishing that x≠0. The error is not necessarily caused by a lack of intelligence. It may be a failure of bookkeeping.

An AI can restate the active constraints at each stage:

We are working under the assumptions that (n) is odd, (p) is prime and (x≠0).

The context becomes a ledger of the reasoning state.

Many difficult mathematical problems contain a profound insight somewhere near the beginning, but they also contain a large amount of less glamorous work afterward: expanding expressions, checking cases, carrying conditions through transformations and making sure that the final conclusion is compatible with every earlier assumption.

A machine does not need to possess deeper insight than a human to gain an advantage here. It may simply be better equipped to preserve the entire state of the problem while completing a long sequence of operations.

Long reasoning chains can exceed human capacity

Mathematics is highly compositional.

A proof can often be represented as:

A → B → C → D

If each step is valid and the chain is preserved accurately, the conclusion follows.

A large working space allows the model to construct much longer chains before losing the thread.

This is important because the difficulty of a problem does not depend only on the difficulty of each individual step. It also depends on how many steps must be coordinated.

A person may be perfectly capable of understanding every local inference in a 100-step argument while still being unable to generate the entire argument unaided. The problem exceeds the person’s ability to maintain the global structure.

AI can potentially compensate by writing down nearly everything.

This may explain why additional “thinking time” often improves model performance. More computation allows the system to produce more intermediate states, examine alternative branches and preserve partial conclusions.

What looks like deeper thought may sometimes be broader search conducted inside a much larger notebook.

Informal reasoning does not work the same way

Now compare a mathematical problem with a social question:

Why has Maria suddenly stopped replying to my messages?

A larger context window might allow an AI to examine years of correspondence. It could identify changes in tone, timing and vocabulary.

But the decisive information may still be missing.

Perhaps Maria is angry. Perhaps she is busy. Perhaps she is ill. Perhaps she has lost her phone. Perhaps she is avoiding an unrelated problem.

No amount of memory can retrieve facts that were never observed.

The challenge is not simply to preserve a known set of premises and derive their consequences. It is to reason under uncertainty about hidden causes.

Informal reasoning also depends heavily on concepts whose meanings are unstable.

Words such as “fair,” “successful,” “responsible,” “harmful” or “intelligent” do not possess the exactness of mathematical variables. Their meaning depends on culture, goals and context.

A model can remember every sentence in a discussion while still misunderstanding what the participants mean.

The same applies to political analysis, historical interpretation, business strategy and psychological judgment. In these domains, the central problem is often not working-memory capacity. It

[truncated for AI cost control]