What Is Working Memory, and Can Training Actually Change It?

This is an article about working memory written by a company that sells a memory game, which is a reason to be sceptical of it. So here is the conclusion first: the evidence that training improves the trained task is decent, the evidence that it improves anything else is genuinely disputed among competent researchers, and we make no claims of the second kind about our app. The rest of this explains what working memory is and why that dispute exists.

The thing you are using right now

Read this sentence, and by the time you reach the end of it you are still holding the beginning. That holding is working memory: the capacity to keep information in an accessible state while you actively do something with it.

Nelson Cowan puts the functional case plainly in his review of the topic — cognitive tasks can only be completed if there is sufficient ability to hold information as it is being processed. Multiplying 47 by 6 in your head requires holding a partial product while you compute the next one. Following an argument requires holding the premises while you evaluate the conclusion. When the holding fails, the task fails, regardless of whether you understood the method.

That makes working memory an obvious target for training, and it is why an entire commercial category exists around it. What follows is an honest account of what is known, what is disputed, and what we will not claim.

How small the capacity actually is

The number most people have heard is seven. In a famous 1956 paper, described by Cowan as humorously complaining about being persecuted by an integer, George Miller reported that people can repeat back a list of no more than about seven randomly ordered meaningful items or chunks.

Later work produced smaller figures. Cowan argues that beneath the task-dependent surface there is a more constant underlying mechanism: a central memory store limited to roughly 3 to 5 meaningful items in young adults. He notes that young adults recall only 3 or 4 longer verbal chunks.

The gap between the two figures is not a contradiction so much as a difference in what is being measured. Performance on a given task can be inflated by strategies — rehearsing, grouping, using long-term knowledge — layered on top of the core store. Strip those away and the underlying capacity looks smaller.

The practical reading of this is uncomfortable and useful: whatever you thought you could juggle, it is probably fewer than that.

The unit is the chunk

The limit is not counted in letters, digits or words. It is counted in chunks — units that have become meaningful as wholes.

The sequence 1 4 9 2 1 7 7 6 2 0 0 1 is twelve digits, which is far past anyone's limit. Read as 1492 1776 2001 it is three chunks, which is comfortably inside it. The digits did not change. The chunking did.

This has two consequences worth carrying around:

Expertise looks like capacity but is not. A chess player who reconstructs a board position after a glance is not holding more than you can. They are holding fewer, larger units, because years of exposure turned configurations into single meaningful things. Their raw capacity is ordinary.

Chunking is trainable in a way that raw capacity may not be. Building the units is a knowledge process, and knowledge accumulates. This is the honest mechanism behind most impressive-looking memory feats, and it is domain-specific by construction — chunks built for chess do nothing for phone numbers.

Near transfer, far transfer, and where the argument is

Two terms make the entire research literature legible.

Near transfer is improvement on the trained task and on tasks that closely resemble it. Practise a digit-span task and get better at digit span, or at a similar span task.

Far transfer is improvement on genuinely different abilities — reasoning, reading comprehension, arithmetic, intelligence test scores, performance at work or school.

Near transfer is comparatively well supported. Far transfer is the disputed territory, and it is also, not coincidentally, what marketing implicitly promises. "Train your brain" is a far-transfer claim wearing a near-transfer claim's clothes.

What the meta-analyses found — including the one that disagrees

This section deliberately presents both sides, because the field genuinely contains both.

Melby-Lervåg and Hulme, 2013, in Developmental Psychology. A meta-analytic review of 23 studies with 30 group comparisons, spanning clinical samples and typically developing children and adults. The programs produced reliable short-term improvements in working memory skills. For verbal working memory those near-transfer effects were not sustained at follow-up; for visuospatial working memory limited evidence suggested they might be. On generalisation, the authors reported no convincing evidence of transfer to other skills — nonverbal and verbal ability, inhibitory processes in attention, word decoding, and arithmetic. Their conclusion was that the programs produce short-term, specific training effects that do not generalise.

Melby-Lervåg, Redick and Hulme, 2016, in Perspectives on Psychological Science. A much larger review: 87 publications with 145 experimental comparisons. Immediately after training there were reliable improvements on measures of intermediate transfer — verbal and visuospatial working memory. On far transfer measures (nonverbal ability, verbal ability, word decoding, reading comprehension, arithmetic) there was no convincing evidence of reliable improvement when working memory training was compared with a treated control condition. Two further findings from that paper matter. First, the degree of improvement on working memory measures was not related to the magnitude of far-transfer effects — which undercuts the proposed causal chain. Second, their publication bias analysis concluded there was no evidential value from the studies using treated controls.

Au and colleagues, 2015, in Psychonomic Bulletin & Review. This is the meta-analysis on the other side, and it should not be left out. Focusing specifically on n-back training and fluid intelligence outcomes, in healthy participants aged 18 to 50 with a control group, the authors included 20 studies and found a small but significant positive effect of n-back training on improving fluid intelligence. They concluded that short-term cognitive training on the order of weeks can produce beneficial effects on important cognitive functions as measured by laboratory tests. The paper attracted published commentary in the same journal the following year — itself a signal of how contested the result was.

Simons and colleagues, 2016, in Psychological Science in the Public Interest. Rather than adding another effect size, this review asks why the field disagrees. Its framing is striking: in 2014 an international group of more than 70 scientists published a consensus statement holding that brain games do not provide a scientifically grounded way to improve cognitive functioning or stave off cognitive decline. Months later a group of 133 scientists and practitioners countered that the literature was replete with demonstrations of benefit. The review sets out and justifies a set of best practices for brain-training interventions, then evaluates the published peer-reviewed intervention studies cited on the websites of leading brain-training companies against those standards.