LLM-Assisted Elaboration Reduces Flashcard Re-Lapse Rates by 48%

1. Introduction

Spaced repetition systems schedule reviews at increasing intervals to maximize long-term retention. When a learner lapses—fails to recall a card—the scheduling algorithm resets the interval, and the cycle restarts. Lapses are costly: each one represents wasted prior study time and signals that the learner’s encoding may be insufficient.

Complement embeds an LLM chat interface directly into the study screen. When a learner encounters a card they’ve forgotten, they can immediately ask the AI to explain the underlying mechanism, provide context, or draw connections to related concepts. The hypothesis is that this interaction creates a richer memory trace—one grounded in understanding rather than rote pattern-matching—that makes the card more resistant to subsequent forgetting.

This is grounded in established cognitive science: elaborative interrogation improves retention over re-reading (Pressley et al., 1987; Dunlosky et al., 2013), generating explanations produces stronger memory traces than passive review (Bjork & Bjork, 2011), and encountering material in different formats creates additional retrieval pathways (Bower, 1972).

2. Data

All data comes from the author’s own account on Complement—the founder’s personal study log spanning August 2024 to March 2026. This is an N = 1 longitudinal study with a large number of within-subject observations.

A lapse is defined as any review with ease ≤ 1 on the 0–4 Anki rating scale. A treatment event is a lapse where an AI chat session occurred within 300 seconds of the review timestamp (median proximity: 37 seconds).

3. Methods

3.1 Analytical designs

We employ three progressively more rigorous designs:

• Raw comparison. All chat-assisted lapses (n = 678) vs. all unassisted lapses (n = 33,501). High-powered but vulnerable to selection bias.
• Stratified matching. Lapse events stratified by previous ease (0–4) and card maturity bucket (reviews 1–3, 4–6, 7–10, 11+). Eliminates confounders correlated with pre-lapse trajectory and card maturity.
• Within-card paired. Cards that lapsed both with and without chat at different times. The same card—same content, same inherent difficulty, same learner—serves as its own control. 472 cards qualify.

3.2 Statistical methods

All confidence intervals are 95% bootstrap CIs (2,000 resamples, percentile method). Between-group comparisons use Fisher’s exact test. Effect sizes are reported as odds ratios with Woolf logit CIs and Cohen’s h for proportions.

4. Results

4.1 Primary finding

4.2 Stratified analysis

After stratifying by previous ease and maturity bucket, the chat effect holds in every stratum where sufficient data exists (n_chat ≥ 5). In 11 qualifying strata, the weighted average re-lapse rate is 3.2% (chat) vs. 6.6% (control)—a 51% relative reduction.

Individual strata lack statistical power (most p > 0.05), which is expected given small per-stratum chat samples. The consistency of direction across all strata is the relevant signal.

4.3 Dose-response

Bucketing chat-assisted lapses by the number of user messages reveals a dose-response relationship:

The optimal engagement is a 3–5 message exchange—enough to ask a question, receive an explanation, and confirm understanding. The slight regression at 6+ messages likely reflects unusually confusing cards requiring extended discussion (note the wide CI).

4.4 Longitudinal trajectory

The chat effect persists beyond the immediate next review. Tracking lapse rates across 5 subsequent reviews:

4.5 Subject area

The stronger effect in Step 1 material (preclinical sciences) may reflect that these subjects have more mechanistic content where an AI explanation adds particular value.

5. Limitations

• N = 1, founder’s data. All data comes from the founder’s personal account. This introduces both a generalizability limitation and a potential conflict of interest. Internal validity is strong (249,281 review events, 472 paired cards), but external validity requires replication across independent users.
• Selection bias. The learner may selectively chat about cards where they feel “close” to remembering. The within-card paired design controls for card-level confounders but cannot fully eliminate within-card temporal selection effects.
• Observational design. This is not a randomized trial. The gold standard would be randomly assigning chats to some lapses and withholding them from others.
• Stratified tests are underpowered. Most individual strata have p > 0.05 due to small per-stratum chat samples. The aggregate pattern (consistent direction across all strata) is the relevant evidence.
• Chat content not analyzed. Not all chats may be substantive. The dose-response relationship provides indirect evidence that quality of engagement matters.

6. Discussion

Across three designs, the re-lapse reduction ranges from 48% to 60%. The most plausible mechanism is that the AI chat helps the learner transition from surface-level memorization to mechanistic understanding. A card previously stored as an isolated fact becomes grounded in a causal model, creating more retrieval pathways and greater resistance to interference.

The magnitude is large by educational intervention standards. For comparison, the testing effect (retrieval practice vs. re-reading) typically produces 20–40% improvement, and elaborative interrogation 15–30%. The AI chat effect may exceed these because it combines elaboration, retrieval, and encoding variability simultaneously at the optimal moment—the point of failure.

The dose-response finding has a direct product implication: the AI should aim for focused, 3–5 turn conversations rather than single-shot explanations or open-ended discussions.

7. Conclusion

Chatting with an LLM after forgetting a flashcard reduces the probability of forgetting it again by approximately 48% in a within-card paired design (OR = 0.46 [0.34, 0.63], p = 2.4 × 10⁻⁷). The effect exhibits a dose-response relationship and persists across 5 subsequent reviews. These findings, while limited to a single user, represent one of the first empirical analyses of LLM-augmented spaced repetition at scale.