Open data
How well do people remember chess positions?
In 3,734 finished rounds of Memory Chess from September 29, 2026 to October 19, 2026, rounds set to 10 seconds of study scored 82% accuracy on average with 2 pieces (380 rounds) and 48% with 6 pieces, the default setting (1,294 rounds). 11.1% of those 6-piece rounds were perfect. With 12 pieces and 8 seconds, a setting fewer players chose, the average was 40%, with 1.4% perfect (74 rounds).
Different players chose each number of pieces, and some rows come from only a few players, so check the Players column in the table before comparing bars.
Different players chose each study time, so these bars do not show how much extra time helps.
Each bar is drawn to scale from 0 to 100% accuracy.
Every published setting
Accuracy by number of pieces and study time. Settings with fewer than 30 rounds or 5 players are left out.
| Pieces | Study time | Rounds | Players | Mean | Median | Middle half | Perfect | Scored 0 |
|---|---|---|---|---|---|---|---|---|
| 2 | 10 s | 380 | 71 | 81.8% | 100% | 100% to 100% | 77.1% | 13.4% |
| 3 | 10 s | 124 | 21 | 85.0% | 100% | 67% to 100% | 72.6% | 4.8% |
| 4 | 10 s | 242 | 40 | 71.0% | 75% | 50% to 100% | 43.4% | 7.4% |
| 5 | 10 s | 101 | 17 | 63.0% | 60% | 40% to 80% | 21.8% | 3.0% |
| 6 | 5 s | 32 | 8 | 54.2% | 67% | 33% to 83% | 9.4% | 18.8% |
| 6 | 10 s | 1,294 | 292 | 48.3% | 50% | 17% to 67% | 11.1% | 15.7% |
| 6 | 15 s | 49 | 6 | 71.1% | 67% | 50% to 100% | 30.6% | 0.0% |
| 6 | 20 s | 113 | 10 | 69.3% | 83% | 50% to 100% | 34.5% | 4.4% |
| 7 | 10 s | 124 | 8 | 61.5% | 71% | 43% to 86% | 12.9% | 2.4% |
| 8 | 10 s | 31 | 6 | 45.4% | 38% | 25% to 75% | 3.2% | 3.2% |
| 12 | 8 s | 74 | 14 | 39.8% | 42% | 17% to 58% | 1.4% | 16.2% |
Method
Every number here comes from the game. When a round's result shows, Memory Chess sends one Google Analytics 4 event, round_complete, with the number of pieces, the study time the round was set to and the round's accuracy. The study time is that setting, not the time the board was actually shown. A player can clear the board early. The event has fired since September 29, 2026, so there is no earlier data. This release covers September 29, 2026 to October 19, 2026, counted in days of the America/Los_Angeles time zone the analytics property uses.
A piece counts only when the right piece of the right colour stands on the right square. Accuracy is the share of the position's pieces placed that way, rounded to a whole percent, minus 10 points for each piece placed beyond the position's count, and never below 0. A perfect round scores 100: every piece right and nothing extra. Most rounds show a new legal position generated at random. A review round replays a board the player saw before, and the daily board is one seeded random position, the same for everyone that day. None is taken from a real game.
A setting is shown only with at least 30 rounds from at least 5 players. Those cuts are my judgement: smaller groups were mostly one or two people, and a setting near the cut is still a small sample. Of 3,734 rounds from 347 players, 2,564 fall in the 11 settings shown, 1,126 in smaller settings, and 44 arrived without their settings. The median and the middle half (the 25th to the 75th percentile) are nearest-rank values of the whole-number scores.
Four settings are one-click presets: 2 pieces at 10 seconds (Easy), 6 at 10 (Medium), 12 at 8 (Hard) and 20 at 5 (Grandmaster). Those rows can include rounds from players who clicked a preset rather than picked a setting. 20 at 5 has too few rounds to show.
The Google Analytics reports behind this release were unsampled and had no “(other)” row, so every setting is counted in full. Google Analytics may withhold or round user counts that fall under its privacy thresholds, so the Players column can be lower than the real number of browsers.
Limits
- Players chose to play and chose their settings, so each row describes the players who chose it, not people in general. The data has no ratings, but in my judgement people who pick harder settings are likely better at this game.
- Repeat players weigh more: one player's 50 rounds count 50 times. Compare the Rounds and Players columns.
- 6 pieces with 10 seconds is the default round, and the home page's calibration round uses the same setting and is pooled into it. It holds the most players (292).
- An empty board counts as a round that scored 0. Since October 14 the game has accepted one empty board a day as a free skip, and the analytics event cannot tell a skip from a real 0, so skips are in the data. The Scored 0 column gives each setting's share, 15.7% for the 6-piece, 10-second rounds.
- Review rounds, which replay a board the player saw before, and the daily board, the same seeded position for everyone that day, are pooled with the other rounds. The data cannot separate them.
- The game changed during this release. The calibration round's event, review rounds, the daily board and the free skip all reached players on October 14, so the 15 days before then hold none of them.
- A player is a browser as Google Analytics counts it. One person on two devices counts twice.
- Rounds from visitors who block analytics, or decline analytics cookies where they are asked, are missing.
- Recall studies find that chess skill helps far more on positions from real games than on random ones (see de Groot's 1944 test, by the numbers). These positions are random, so in my judgement they measure short-term memory for pieces more than chess knowledge.
How to cite
Cheng, B. (2026). Memory Chess position recall data, September 29, 2026 to October 19, 2026 [Data set]. Memory Chess. https://thememorychess.com/data
The data is free to reuse under CC BY 4.0: credit Memory Chess and link this page. The CSV and the field notes are on GitHub, with a new release planned each month.