The sentence you practised yesterday, the one that felt smooth by the third repetition, is gone this morning.
Almost everyone who practises speaking seriously hits this wall, and most people read it as "bad memory" or "I didn't concentrate enough."
Neither. This is what memory does, and it was measured more than a century ago.
The forgetting curve, 1885
Hermann Ebbinghaus did something crude and effective: he memorised material, tested himself after varying delays, and plotted how much survived. The result is the forgetting curve.
Its shape: steepest loss right after learning, flattening out over time.
Two consequences:
1. Losing most of it by the next day is the normal value, not a failure. 2. What determines how much you keep isn't the session where you learned it — it's how many times you come back before the curve bottoms out.
The spacing effect: same reps, better placed
Ebbinghaus also found something more useful, now called the spacing effect: the same number of repetitions, spread over days, produces much better retention than the same repetitions crammed into one session.
This isn't a small-sample result. In 2006, Cepeda and colleagues published a meta-analysis covering 184 articles, 317 experiments and 839 separate assessments. The conclusion was unambiguous: distributed practice reliably beats massed practice.
One detail from that meta-analysis gets overlooked: the best gap depends on how long you need to remember. For a week, about a day. For a year, several weeks. The further out the target, the longer the interval.
For speaking, there's a crucial difference
Most people meet spaced repetition through vocabulary flashcards. Speaking is different, and the difference matters:
Flashcards rehearse recognition. Speaking requires production.
Seeing a card and recalling the meaning, and opening your mouth and producing the sentence, run on different pathways. So:
Reviewing a sentence for speaking means saying it out loud again. Reading it again doesn't count.
Re-reading the phrase, glancing over your notes — it feels like review, and it trains recognition. The only form of review that works for speaking is producing it.
A schedule you'll actually keep
No algorithm needed. Four touches:
- Day 0: say it 3 times (this is building, not reviewing)
- Day 1: say it once ← the important one, landing where the curve is steepest
- Day 3: once
- Day 7: once
Five spoken repetitions per sentence, spread across a week, under two minutes total. Far more effective than ten repetitions on day one — which is precisely the spacing effect: not more reps, better placed reps.
The part people get wrong: what goes in the queue
This is where spaced repetition usually collapses. The queue swells, review becomes a chore, and the whole thing stops.
The filter is simple:
- Only sentences you actually stumbled on. The ones that came out fine don't need a slot.
- One entry per problem. You may have made the same mistake in five sentences; keep one representative.
- Allow deletion. Once something is automatic, take it out. The queue should shrink sometimes, not only grow.
Queue length isn't a measure of diligence. It's a measure of debt.
How MirrorSay handles it
MirrorSay's review queue uses exactly that filter: a sentence goes in only if you practised it and didn't keep up at the time. Repeats of the same sentence merge instead of taking new slots, and once it's automatic you can remove it.
And review means saying it again, not reading it — which, per the above, is the only form that counts for speaking.
Related: Most of What Natives Say Is Prefabricated explains why recognition and production are separate systems in the first place; How to Practice English Speaking Alone has the session itself.
References
- Ebbinghaus, H. (1885). *Über das Gedächtnis* (Memory: A Contribution to Experimental Psychology). First systematic measurement of the forgetting curve and the spacing effect.
- Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks: A review and quantitative synthesis. *Psychological Bulletin*, 132(3), 354–380. (184 articles, 317 experiments, 839 assessments)
*MirrorSay is a free AI speaking-practice app: say a sentence, see how a native would say it, then shadow it. No sign-up, and your voice stays on your phone.*