Quantum Country users rarely forget after demonstrating five-day retention

In terms of accuracy rates

Quantum Country readers rarely forget answers to questions they’ve been able to remember for 5+ days: the median user’s accuracy is 95% from that point on. And it climbs as higher intervals are demonstrated.

Per-user accuracy rates in traces after reaching these levels of demonstrated retention (25 / 50 / 75th %ile) 20210316162015:

  • 5 days: 89% / 95% / 98%
  • 2 weeks: 89% / 95% / 98%
  • 1 month: 91% / 96% / 99%
  • 2 months: 93% / 97% / 99%
  • 4 months: 96% / 99% / 100%

These are calculated over traces of at least a year in total length. The numbers don’t vary all that much with different lengths.

In terms of total repetitions

For the 53 readers who have reviewed every QCVC card for a full year following their first successful recall, the median (IQR) review count is 463 (436-507) and lapse count is 10 (22-38). So for most readers, less than 2% of repetitions constitute a lapse. That rises to 7% for 25th %ile readers. 20211126101825

For half a year, the same results (N=129): 332 (304-358) reviews; 14 (4-26) lapses. (Funny that going from half a year to a year involves only ~1/3 more repetitions)

In terms of “perfect” memory traces

In fact, the vast majority of reader-item traces contain zero lapses after they’ve been remembered:

  • after demonstrating 5 days retention: 82% of reader–item traces contain zero lapses
  • … 2 weeks: 82%
  • … 1 month: 89%
  • … 2 month: 93%
  • … 4 month: 97%

20210316161406 (again, over 1 year in length)

This is of course a function of our schedule. The differences here may be more about repetition count than demonstrated retention interval.

Slicing by question

See 95% of Quantum Country questions produce no lapses for most readers’ first half year

Slicing by user

How many misses-per-question do readers have in their first half year?

How many lapses-per-question do readers have in their first half year?

The distribution of lapses-per-question by user, over the first half year (5% outliers in final bucket):


This isn’t really a power-law like we see for questions (e.g. Half of all long-term Quantum Country lapses come from just 12% of its questions). The CDF increases fairly uniformly over ~80% of users. The last two deciles have about half the slope.

Looking at all users who collected every card from QCVC (N=107), absolute lapse counts 20210325122332:

This is pretty close to two linear paths. Pretty steady for 70-80% of users. The median (IQR) user has 11 (5-20) total lapses. This needs to be squared with 95% of Quantum Country questions produce no lapses for most readers’ first half year; it suggests that users are lapsing on different questions. For reference, the median (IQR) number of reviews users engaged with during these periods is 342 (334-354), so the median user has an overall accuracy rate of 97%. Extending the timeframe to one year, the median (IQR) number of lapses is 15 (6-35), and the median (IQR) number of reviews is 459 (446-483), across 48 readers.

What about the number of forgotten attempts recorded in histories before hitting five days of demonstrated retention? 20210325142719. The median (IQR) reader has 8 (5-19) lapses. So for most people, there really is a modest amount of churn happening in the first few review sessions. Also: Quantum Country readers forget about as many times before their first successful repetition as they do in the following half year.

Related, a more specific look at higher-level accuracy: Demonstrated retention reliably bounds future recall attempts on Quantum Country

Caveats: Recall rates are a misleading proxy for more meaningful goals of the mnemonic medium; Quantum Country’s high recall accuracy rates may mask forgetting because of cuing

Last updated 2023-07-13.