What Adaptive Language Apps Should Track

The short answer

An adaptive language app should track three things in a closed loop: which concepts are due (vocabulary and grammar items scheduled by spaced repetition), what exercise was generated for those due items, and word- or concept-level outcomes from the learner's attempt. Without all three, "adaptive" usually means a level estimate, a streak, or a fixed deck, not a schedule that updates from what you actually retrieved.

Retrieval practice strengthens memory when the system can attribute success or failure to specific items, not only to a whole lesson.¹ Modern schedulers such as FSRS need clean per-item grades to predict when recall will fail next.² LinGoat is one example of combining due-concept scheduling, novel sentence generation, and granular grading; the checklist below is the capability map, not a product tour.

Why "adaptive" needs more than a level badge

Many apps call themselves adaptive because they skip easy units, raise difficulty after a quiz, or estimate a CEFR band. That can personalize content selection. It does not, by itself, tell the app which conjugations or lemmas are about to be forgotten tonight.

Spaced repetition works by interrupting the forgetting curve at the item level: each word or grammar concept has its own history of retrieval attempts.³ If the app only stores "Lesson 12: 80%," it cannot reschedule the one preposition you missed while advancing four solid nouns. Adaptive practice for productive language therefore has to log memory units, exercise context, and granular outcomes together. For the memory model behind timing, see how spaced repetition works.

1. Which concepts are due (SRS)

The first signal is the due queue: which tracked concepts should be practiced today so predicted recall stays near a retention target. Older SM-2-style systems mostly stretch intervals with an ease factor. FSRS models difficulty, stability, and retrievability per item and places the next review when recall probability is expected to hit the target.²

Large Anki-collection benchmarks show FSRS predicting recall more accurately than SM-2 on card-level histories (treat published RMSE and superiority figures as snapshot evidence, not eternal constants).⁴ For language learners, that accuracy matters because cognates and irregular verbs do not forget on the same curve. An adaptive app that "knows your level" but cannot say which five items are due is not scheduling; it is sequencing content. Deeper comparison: FSRS vs SM-2 for language learning.

2. What exercise was generated

The second signal is the exercise shell: what prompt the learner saw for those due concepts. Tracking the exercise matters for two reasons.

• Auditability: You can tell whether a failure came from a fair retrieval attempt or from a broken prompt, an unnatural pile-up of items, or a recognition task that never forced production.
• Transfer: If the same static sentence is reused, learners can pass by recognizing the string rather than retrieving the components. Novel sentences that still pack due items keep practice closer to real composition.

For production-oriented apps, a strong default is native-to-target sentence translation that packs as many due concepts as possible while staying natural. The sentence is the delivery format; the due list is the curriculum for that moment. How packing and scheduling fit together is covered in multi-concept sentence reviews.

3. Word- and concept-level outcomes

The third signal is granular grading: each vocabulary and grammar concept in the attempt scored on its own, then written back to the scheduler. Whole-sentence pass/fail contaminates the log. Miss one ending and four solid recalls get punished; nail most of the sentence and a weak form rides along as "Good." Either way, FSRS (or any per-item model) trains on noisy labels.

Testing-effect research treats the retrieval event itself as the learning mechanism.¹³ For sentence practice to feed that mechanism, the app must resolve which concepts were retrieved. That is what LinGoat calls granular attribution: word-by-word (and grammar-by-grammar) outcomes inside a full production attempt. See word-by-word grading for language learning.

Capability checklist

Use this as a buyer or builder checklist. An adaptive production loop is credible when the product can show:

• Concept inventory: Vocabulary and grammar tracked as separate scheduleable items, not only as lesson scores.
• Due queue: A daily (or session) set of concepts selected by spaced repetition, ideally with a calibrated model such as FSRS rather than fixed "review this chapter" blocks.
• Exercise generation tied to dues: Prompts built from the current due set (often novel multi-concept sentences), not a random shuffle of static cards.
• Production requirement: Typed or written recall without multiple-choice lures, so the attempt is a genuine retrieval event.
• Per-concept grades: Success and failure attributed to each item in the answer, with immediate localized feedback.
• Scheduler write-back: Each grade updates that item's next interval; known pieces are not dragged down by a single typo.
• Honest analytics: Progress reported as active mastery of concepts (stability, due load, vocabulary size), not only streaks or XP.

Missing any link breaks the loop. SRS without generation invites static memorization. Generation without granular grades corrupts the schedule. Grades without a due queue become a one-off quiz.

Where many apps stop short

Common gaps, even in polished products:

• Recognition graded as mastery: Multiple-choice or tap-to-match logs feel like data but mainly measure recognition, which systematically overestimates productive recall.
• Sentence cards without item grades: Whole-string reviews look efficient until one error resets five intervals.
• Level engines without item histories: Placement tests adapt content bands; they do not maintain forgetting curves per lemma.
• Gamification metrics as substitutes: Streaks and XP can support persistence, but they are not evidence that the right concepts returned at the right time.

None of these make an app useless. They do mean the marketing word "adaptive" is doing more work than the data model.

LinGoat as an example of the combination

LinGoat implements the three-signal loop for written sentence practice: FSRS decides which concepts are due; the generator packs those dues into novel native-to-target translation prompts; granular attribution grades each concept and reschedules it. The app is not a speaking replacement, and dedicated speaking modes that reuse the same mastery map are roadmap territory, not a claim about what ships today.

For the full pedagogical frame (production vs recognition, laddering, and why the three steps must travel together), read LinGoat's full pedagogy.

See how LinGoat works or try the app.

References

1. Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249-255. https://doi.org/10.1111/j.1467-9280.2006.01693.x
2. Ye, J., Su, J., & Cao, Y. (2022). A stochastic shortest path algorithm for optimizing spaced repetition scheduling. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 4384-4394). Association for Computing Machinery. https://doi.org/10.1145/3534678.3539081
3. Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966-968. https://doi.org/10.1126/science.1152408
4. Expertium. (2025). Benchmark of spaced repetition algorithms. https://expertium.github.io/Benchmark.html