Habitual vs. Episodic Retention: What the Base Rates Actually Are

The widely quoted 254-day upper bound for habit formation is an extrapolation beyond what Lally et al. (2010) actually measured, not a reported finding. Separately, Wood & Neal (2007, Psychological Review) establish that habit formation requires an externally supplied, recurring cue — nothing in a self-directed learner's daily life supplies a natural cue to "study today," unlike, for example, brushing teeth, which piggybacks on an existing routine.

Kerfoot's subjects were practicing medical professionals engaging with career-relevant content delivered directly into their inbox — an external, professional stake that a general self-directed learner does not have. The design principle (bring the actual content to the learner rather than requiring them to return to a destination) is reasoned to transfer; the specific 74–77% completion magnitude is not expected to.

Ghibellini & Meier (2025, Humanities and Social Sciences Communications) meta-analyzed the Zeigarnik effect — the claim that interrupted tasks are recalled better than completed ones — and found a weighted interrupted-to-completed recall ratio of 0.99, indistinguishable from chance; interrupted tasks made up 49.16% of recalled tasks. The authors conclude the Zeigarnik effect "lacks universal validity." A related but distinct effect — the Ovsiankina effect, the tendency to want to resume a concretely unfinished task — does replicate in the same 2025 paper. The distinction matters: manufactured narrative suspense is not supported; a genuinely unfinished, resumable task is.

The primary peer-reviewed evidence is Silverman & Barasch (2023, Journal of Consumer Research 49(6), 7 studies), which shows that displayingan intact versus a broken streak changes engagement while holding actual behavior constant — a framing effect on an existing log, not direct evidence that adding a streak mechanic raises retention. The same paper finds streak breaks are amplified by self-blame and attenuated when the streak is repairable. Commonly circulated numbers such as "streaks lifted retention 12%→55%" or "Streak Freeze cut churn 21%" are marketing claims with no published method behind them. Duolingo's own peer-reviewed retention work (Yancey & Settles, KDD 2020) — a bandit algorithm for notification timing — produced +0.5% DAU and +2% new-user retention, which is the actual documented magnitude of best-in-class execution at the most-resourced company in this category.

Gal, Rucker & Shavitt (2018, Journal of Consumer Psychology28) argue the current evidence does not support the claim that losses generally loom larger than equivalent gains. "Losses hurt twice as much, therefore streak mechanics work" is a chain of reasoning that rests on a premise the literature no longer treats as settled.

Bell et al. (2023, JMIR mHealth 11:e38342), a 30-day micro-randomized trial with 566 users, found notifications produce a 3.5× lift in the probability of opening the app within one hour (95% CI 2.91–4.25), decaying to a 1.3× lift at 24 hours — but found no significant difference in time-to-disengagement across aggressive, random, and zero-notification policies (median 11 vs. 11 vs. 7 days). Notifications move when people show up in the short term; the trial found no evidence they change whether a user ultimately disengages.

Sailer & Homner (2020, Educational Psychology Review 32) report cognitive outcomes at g = 0.49 and motivational outcomes at g = 0.36, but behavioral outcomes — the ones that define retention — at only g = 0.25, 95% CI [0.04, 0.46], k = 9: the fewest studies, the widest confidence interval, and an effect that did not survive the high-methodological- rigor subsplit in the same analysis. Hanus & Fox (2015, Computers & Education 80), a quasi-experimental (not randomized) comparison, found a badge-and-leaderboard course section declined in intrinsic motivation, satisfaction, and empowerment, and scored lower on the final exam than the non-gamified section.

Lally et al.'s 66-day median is evidence against trying to install daily-habit mechanics for a learner with no external stake — it describes how long habit formation takes even under favorable conditions with a pre-existing cue, not a recipe for achieving it in a general product.