First Forgetting Curve in Coercive Learning

Analysis of retention patterns using Algorithm SM-17

Dataset Overview

This analysis examines retention data collected using the SM-17 spaced repetition algorithm. The dataset contains 35 data points spanning intervals from 1 to 2,296 days, with recall rates ranging from 0.34 to 0.73. Unlike classical forgetting curves, this data shows the complex retention patterns characteristic of optimized spaced repetition systems.

Power Law Model

R(t) = a × t-b + c

a (amplitude): 0.1456
b (decay rate): 0.0423
c (asymptote): 0.3500
R²: 0.0294

Exponential Model

R(t) = a × e-bt + c

a (amplitude): 0.1929
b (decay constant): 0.000110
c (asymptote): 0.3500
R²: 0.0250

Model Comparison

Better Fit: Power Law
Difference: 0.0044
Data Points: 35
Time Range: 1-2,296 days

Analysis Summary

Data Characteristics: The dataset exhibits significant variability typical of real-world spaced repetition learning.

Power Law Performance (R² = 0.0294): The power law model shows a slightly better fit and is theoretically more appropriate for heterogeneous spaced repetition data. The small decay exponent (b = 0.0423) indicates very gradual forgetting.

Exponential Performance (R² = 0.0250): The exponential model, while showing similar performance, has an extremely small decay constant (b = 0.000110), suggesting minimal forgetting over time.

Key Insights:

  • Both models converge to a retention baseline of ~35%, indicating stable long-term memory consolidation
  • High variability suggests individual differences and item-specific factors play major roles
  • Power law behavior aligns with research on memory consolidation and learning curves in heterogeneous coercive condition