In this study we examined children’s performance in an educational web game. We published a new game containing 10 different working memory intensive tasks. The presentation modality was visual in four games and auditory in six games. The difficulty level of the game – the amount of given stimulus in the level – adapted to player skills with bayesian algorithm. Ability to understand Finnish and numbers was needed to successfully play the game. We advertised the game in kindergartens and in a Facebook group for mothers of children with special needs. We got 1 074 volunteer players’ who performed 23 315 game tries, average being 22 and median 8 tries per player. Game results got statistically better with playing but only within the limits of players memory capacity. There was very small change in maximum results in different game levels and the noticable change happened in average results. In addition, we divided the players into two groups, one of which was given 30 % chunked stimulus. There was no improvement effect between groups, and even worse, on the memory span of two results were interfered by chunking. Finally, in our study we reproduced, with statistically higly significant results (N = 443, p < 0,001), a well known phenomenon called prior-list intrusion, where stimulus from the previous trials affect responses of the subsequent trials.