feat: collaborative interrogation (how/why deepening) + Ebbinghaus forgetting curve visualization #76
Description
Summary
Two complementary learning science features:
- Elaborative interrogation — during quiz sessions, prompt the learner with "how?" and "why?" follow-up questions to deepen understanding beyond surface recall
- Forgetting curve visualization — show the Ebbinghaus exponential decay curve for concepts, making memory decay tangible and motivating timely review
Elaborative Interrogation
Elaborative interrogation is an evidence-based learning technique where learners are asked to explain how and why something works, rather than just what it is. This forces deeper processing and stronger memory encoding.
How it could work in Engram
- After a quiz answer, follow up with a generated "why?" or "how?" question (e.g., after recalling "CRDT stands for Conflict-free Replicated Data Type" → "Why does conflict-free replication matter for local-first architectures?")
- Claude generates contextual follow-ups based on the concept's relationships in the knowledge graph
- Correct deep answers could boost FSRS difficulty/stability more than surface recall alone
- Collaborative = the AI and learner explore understanding together, not just testing rote memory
Forgetting Curve Visualization
Ebbinghaus's forgetting curve shows that memory decays exponentially over time without reinforcement. Visualizing this makes the abstract concrete.
How it could work in Engram
- Show per-concept or aggregate forgetting curves on the dashboard or concept detail view
- Plot predicted retention (from FSRS retrievability R) over time as a decaying exponential
- Highlight where review is scheduled vs. where the curve drops below desired retention
- Animate the curve "resetting" after each successful review — showing how spaced repetition fights the decay
- Could overlay multiple concepts to show which ones are decaying fastest
Research: Elaborative Interrogation
Foundational Research
Pressley et al. (1988) conducted the seminal study. Subjects constructed reasons why each presented fact "made sense." Memory was consistently and substantially better in the elaborative-interrogation condition than in reading-control conditions, particularly for "confusing facts" — arbitrary pairings difficult to remember without elaboration.
Pressley, M., Symons, S., McDaniel, M. A., Snyder, B. L., & Turnure, J. E. (1988). Elaborative interrogation facilitates acquisition of confusing facts. Journal of Educational Psychology, 80(3), 268-278.
Woloshyn, Pressley, & Schneider (1992) demonstrated that the technique's effectiveness is modulated by prior knowledge: learners with existing domain familiarity benefit significantly more, because they can generate richer explanations. This aligns with Engram's prerequisite-based unlocking — interrogation questions should only appear for concepts whose prerequisites are mastered.
Dunlosky et al. (2013) Meta-Analysis
The landmark review of ten learning techniques rated elaborative interrogation as moderate utility:
| Utility Rating | Techniques |
|---|---|
| High | Practice testing, Distributed (spaced) practice |
| Moderate | Elaborative interrogation, Self-explanation, Interleaved practice |
| Low | Summarization, Highlighting, Keyword mnemonic, Imagery for text, Re-reading |
Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4-58.
Effect Sizes
Donoghue & Hattie (2021) meta-analyzed all ten Dunlosky techniques across 242 studies (169,179 participants, 1,619 effect sizes):
| Technique | Effect Size (d) |
|---|---|
| Practice testing | 0.74 |
| Elaborative interrogation | 0.56 |
| Distributed practice | High |
The effect is strongest when elaborations are self-generated (not provided), precise (not vague), and learners have sufficient prior knowledge.
Donoghue, G. M., & Hattie, J. A. C. (2021). A meta-analysis of ten learning techniques. Frontiers in Education, 6, 581216.
Synergy with Spaced Repetition
The Learning Scientists (Weinstein et al., 2018) explicitly recommend combining elaborative interrogation with retrieval practice: "Ideally, students would be able to describe and explain ideas from memory." The combination — retrieving from memory, spacing that retrieval over time, and generating "why/how" explanations at each retrieval — represents a particularly potent learning sequence.
Brown, P. C., Roediger, H. L., III, & McDaniel, M. A. (2014). Make it stick: The science of successful learning. Harvard University Press.
AI-Assisted Socratic Questioning
Socratic Chatbot (Mavriki et al., 2024): A Llama2-based Socratic tutor significantly outperformed non-Socratic baselines (p < 0.001). Students showed progressively improved critical thinking over 5 conversation turns. Lower-performing students benefited disproportionately.
Key finding: Consistent exposure to AI-guided questioning develops metacognitive fluency — students learn to anticipate the next question, internalizing the Socratic process as self-dialogue.
Mavriki, P., et al. (2024). Enhancing critical thinking in education by means of a Socratic chatbot. arXiv:2409.05511.
Research: Ebbinghaus Forgetting Curve
Original Work (1885)
Hermann Ebbinghaus published Über das Gedächtnis (1885), establishing the first quantitative study of human memory using nonsense syllables. Key findings: memory decays rapidly at first then levels off (roughly 50-60% lost within the first hour), the rate is not linear but exponential, and relearning is faster than initial learning ("savings").
Modern Replication: Murre & Dros (2015)
Murre and Dros published a landmark replication in PLOS ONE with a subject spending 70 hours learning and relearning 70 lists of 104 nonsense syllables. Both the power function and Memory Chain Model achieved excellent fits (R² = 0.987 and 0.996 respectively). Despite being conducted over a century later, the replication curves closely matched Ebbinghaus's originals. They also found an unexpected 24-hour "bump" in savings scores, possibly from sleep consolidation.
Murre, J. M. J., & Dros, J. (2015). Replication and analysis of Ebbinghaus' forgetting curve. PLOS ONE, 10(7), e0120644.
The Mathematical Model
The modern exponential decay model:
R = e^(-t/S)
where R = retrievability, t = time elapsed, S = stability. In practice the forgetting curve is better described by a power function (aggregate of traces with different stabilities):
R(t) = (1 + F · t/S)^C
Wozniak's Two-Component Model (1995) formalized the distinction between Stability (storage strength) and Retrievability (retrieval strength). Critical insight: repetitions have minimal power to increase stability when retrievability is still high — this is the spacing effect. This model provides the theoretical foundation for all modern spaced repetition algorithms.
How Spaced Repetition Counteracts the Curve
Each review just before forgetting: (1) restores retrievability to ~100%, and (2) increases stability so the next curve decays more slowly. This produces a "sawtooth" pattern.
Cepeda et al. (2006): Meta-analysis of 839 assessments across 317 experiments found the optimal inter-study interval is approximately 10-20% of the desired retention interval (want to remember for 1 year → review every 5-10 weeks).
Cepeda, N. J., Pashler, H., Vul, E., Wixted, J. T., & Rohrer, D. (2006). Distributed practice in verbal recall tasks. Review of General Psychology, 10(4), 354-380.
FSRS: State of the Art
FSRS achieves a 99.6% superiority rate over SM-2 on log-loss (349M+ reviews from 10K collections). Key advances: power-function forgetting curve, stability saturation, retrievability-aware scheduling, mean-reverting difficulty (preventing "ease hell"), and per-user parameter optimization via machine learning.
Visualizing the Curve as a Learning Tool
Blech & Gaschler (2018) found that students' subjective beliefs about forgetting often deviate significantly from the actual curve shape (expecting linear rather than exponential decay). Dysfunctional metacognitive beliefs lead to suboptimal study strategies. Interactive demonstrations of forgetting curves help consolidate accurate metacognitive knowledge.
Blech, C., & Gaschler, R. (2018). Assessing students' knowledge about learning and forgetting curves. Psychology Learning & Teaching, 17(3), 308-325.
Implication for Engram: Showing personal forgetting curves (predicted vs. actual retention over time) serves as a metacognitive calibration tool. The visual "sawtooth" pattern — where each review resets retrievability but with progressively shallower decay — is particularly powerful for building intuition about why spacing works and cramming doesn't.
Key Milestones
| Year | Researcher(s) | Contribution |
|---|---|---|
| 1885 | Ebbinghaus | First quantitative forgetting curve |
| 1967 | Pimsleur | Graduated interval recall schedule |
| 1992 | Bjork & Bjork | New Theory of Disuse (storage vs. retrieval strength) |
| 1995 | Wozniak et al. | Two-component model (stability + retrievability) |
| 2006 | Cepeda et al. | Meta-analysis: optimal ISI is ~10-20% of retention interval |
| 2015 | Murre & Dros | Successful replication of Ebbinghaus |
| 2022-2026 | Ye (Expertium) | FSRS: ML-optimized DSR model |
Synergy
These two features reinforce each other: the forgetting curve shows when you need to review, elaborative interrogation improves how deeply that review encodes. Together they shift the app from "flashcard driller" to "understanding partner."
Related
- FSRS migration (
docs/FSRS_MIGRATION.md) — retrievability R is literally the forgetting curve; FSRS already models it, we just need to visualize it - Feature: Claude-computed concept embeddings for semantic discovery #39 — Concept embeddings (richer context for generating interrogation questions)
- feat: video-synchronized knowledge graph highlighting with relationship explanations #74 — Video-synced highlighting (interrogation could happen during video pause points too)