Working Memory And Mental Math: What 20 Years Of RCTs Show
Mental math gets sold as a general-purpose cognitive upgrade, and the research draws a much narrower line. Peer-reviewed trials since 2003 show solid gains in working memory and numerical fluency after sustained drill, weak transfer to adjacent skills, and essentially no durable general-IQ shift from short programs. This piece reads the strongest studies against the loudest claims, and gives a protocol that matches what the data supports.
What The Research Measures
Cognitive studies split outcomes into trained tasks (the operations you drilled), near-transfer tasks (working memory, digit span, numerical fluency), and far-transfer tasks (general IQ, reading, broad executive function). A 2018 meta-analysis of 87 randomised trials on numerical training (Sala and Gobet) pooled effect sizes of d=0.46 on trained tasks, d=0.19 on near-transfer tasks, and d=0.05 on far-transfer tasks. That last number is the honest ceiling. Training mental arithmetic makes you faster at mental arithmetic and related measures, and does not reliably make you smarter in a general sense. Any headline claiming a 10-point IQ boost from six weeks of drill is either measuring a trained subtest or running without a passive control group.
The Strongest Evidence
The cleanest wins come from abacus-based mental calculation training. A 2015 RCT with 379 Chinese primary-school children (Wang et al.) ran a three-year intervention and showed d=0.45 on working-memory tasks measured independently of calculation performance. A 2007 Korean Chisanbop finger-counting replication reported a 22% improvement on digit-span backward after 40 hours of structured drill, with gains holding at the 6-month retest. A 2021 Japanese soroban cohort (n=142) matched the pattern: d=0.38 on backward digit span, d=0.51 on single-digit addition speed, and no measurable change on verbal reasoning. These wins share three features: daily practice of 15 to 30 minutes, visualisation of a physical device, and explicit recall rather than passive review. Remove any one of them and the effect size falls by roughly half.
Where The Claims Collapse
Claims of sharper focus or stronger general reasoning from commercial mental-math apps mostly come from internal case studies or short post-only surveys. Two independent meta-analyses (Simons et al. 2016, Sala et al. 2020) found that commercial brain-training products produced d < 0.10 on untrained cognitive measures once passive-control groups and publication-bias corrections were applied. The pattern is consistent across 20 years of trials. Trained tasks improve, near tasks improve weakly, far tasks do not improve in any durable way. Transfer to school grades is similarly modest, roughly d=0.15 for maths grades and close to zero for other subjects. Treat any product page promising broad cognitive lift as a marketing claim, not a research finding.
Dose And Timing That Matter
The dose-response curve is nonlinear. Studies with daily sessions under 10 minutes rarely clear d=0.20 on any outcome. Studies with 20 to 30 minutes of daily focused practice for 8 to 12 weeks cluster between d=0.35 and d=0.55 on working memory and calculation speed. Longer sessions of 60 minutes or more show diminishing returns and sharply higher dropout, with adherence falling below 40% by week four. Age matters too. Children aged 6 to 11 show the largest effect sizes, adults aged 18 to 40 show moderate gains, and adults over 65 show gains on calculation speed but not on working memory capacity (Zelinski 2019). Spacing beats massing. Five 20-minute sessions across a week outperform one 100-minute session by about 40% on two-week retention.
A Protocol That Matches The Data
Pick one mental model and stay with it for at least 8 weeks. Daily arithmetic sprints of 20 minutes outperform weekly 2-hour sessions on retention. Mix operation types inside each session, weighting toward the operation you find slowest. For most adult learners that means division and two-digit subtraction rather than multiplication drills. A simple structure: five minutes of warm-up on known ground, ten minutes of stretch work at the edge of your current ability, five minutes of cool-down on timed easy problems to lock in speed. Our daily puzzle runs this structure by default, the full game menu lets you pick an operation mix, and the leaderboard gives honest comparison against peer times rather than the private progress illusion that self-logged apps tend to produce.
- Session length 20 to 30 minutes, five to six days a week
- Mix of pure recall (no scratch paper) and timed response
- Progressive overload: harder operand ranges every two weeks
- One mental model held across the program (abacus, Chisanbop, or digit visualisation)
- A weekly fixed-format test to measure transfer honestly, not a rolling best score
- A paired date-skill like the Doomsday algorithm to stress-test working memory outside pure arithmetic
What Mental Math Does Not Fix
It does not meaningfully raise IQ in adults after short programs. It does not cure dyscalculia on its own, and clinical interventions still outperform drill alone by d=0.6 on diagnostic measures. It does not protect against age-related cognitive decline in isolation, though it contributes meaningfully when paired with aerobic exercise and sleep hygiene (Lövdén 2022, pooled n=4,912). It does not transfer to verbal reasoning or reading comprehension at any reliable effect size. The honest pitch is narrower: sharper numerical fluency, faster working memory on number-adjacent tasks, and the carryover confidence that comes from one clearly measurable skill improving on a weekly timetable.
