Original title: Evaluation of Memory Decay Models in Spaced Repetition Data
Vedh Kannan, Niyati Ravi · September 2026
Scientists still argue over how human memory fades over time. Many think forgetting follows a power law, where you forget quickly at first and then much slower. However, combining different people's data can create an artificial curve that mimics a power law. Researchers studied review logs from learning apps Anki and Duolingo to test exponential, power, and stretched exponentialA pattern that starts fast but slows more than a regular exponential. curves. These apps use adaptive algorithms that schedule reviews based on estimated memory strength, linking review delays directly to memory. The researchers found this scheduling changes which items get tested as intervals grow, creating a false power-law pattern. Simple averaging cannot explain the curve and contradicts real study histories. Looking at individuals in the Anki data, 76% of cases actually followed a stretched exponential decayA pattern where forgetting happens at a constant rate over time.. Therefore, studies using adaptive apps must account for how the scheduler works to determine true forgetting curves.
Imagine you want to measure how quickly coffee cools. You check the temperature every few minutes, but you decide when to check based on how hot you think it is: hot coffee gets checked often, cooler coffee less often. Your measurements will show a cooling pattern that depends on your checking schedule, not just the coffee's natural cooling. Here, the researchers found that spaced repetition apps schedule reviews based on predicted memory strength, which distorts the observed forgetting curve in the same way.
Anyone using spaced repetition apps (like Anki or Duolingo) depends on algorithms that schedule reviews. If those algorithms assume the wrong memory decay model, they may review items too early or too late, wasting time or causing forgetting. This research shows that the forgetting curve observed in these apps is distorted by the scheduling itself, and that the true shape is likely stretched exponential. Better memory models require accounting for this scheduling distortion.
The abstract does not establish the true forgetting function outside the two platforms studied. It only tests three specific models (exponential, power, stretched exponential) and does not propose a new model. The individual-level analysis only covers Anki data, and 24% of users did not fit the stretched exponential. The study is observational, not controlled, so causal conclusions about the scheduler's effect are limited. Additionally, it does not provide a method to correct for the artifact, only recommends modeling the scheduler's belief state.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Evaluation of Memory Decay Models in Spaced Repetition Data — Vedh Kannan, Niyati Ravi — EdArXiv — doi:10.35542/osf.io/5hqc9_v1
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.