What does Mid-Context Forgetting and Relay Thresholds conclude, in numbers, about “The Classics Examined: Mencius on One Day of Sun and Ten Days of Cold, and Ebbinghaus on Memory”?

Answer: Program p72 turns Mencius's image into parameters: with a five per cent cold-day loss, the warm day must add 0.670 to break even; in a relay, restating a key fact only every eleventh layer leaves a mean survival of 0.863 and a worst of 0.737. Program p71 shows that Ebbinghaus's original data look more like a power law (an error of 0.0206 against 0.0933) while the replication looks more like an exponential, so the form of forgetting is not settled. Both classics point to the same business conclusion: what is key must get sun every day, and strong and weak facts must be told apart and counted separately. The last chapter must say what workflow an owner can follow tomorrow morning, and what result would overturn this book.

Scope of the numbers: Tasks, material and parameters are illustrative and the model constants come from one 2023 model; all calculations are generated by the book's programs, which call no model. It is not individual business advice and neither rates nor recommends any model or service.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. printed values that match the answer: 0.670, 0.737, 0.863

cold-day loss ccycle growth g=0.3, w=10break-even g*
0.011.1760.106
0.030.9590.356
0.050.7780.670
0.100.4531.868
restate every m layers at r = 0.97:
mmean retentionworst
11.0001.000
20.9850.970
40.9560.913
110.8630.737
Show the full program
# Chapter 7. p72: "one day of sun, ten days of cold" as a parameter.
# Part 1 (Mencius, Gaozi I, plain reading): a seedling gains a share g
# on a warm day and loses a share c on each cold day; over a cycle of
# one warm and w cold days it grows by (1 + g)(1 - c)^w. Break-even
# g* = (1 - c)^(-w) - 1. Part 2 (relay): a key fact is restored from
# the source every m layers and decays by r per layer in between; at a
# random layer its mean retention is (1 - r^m) / (m (1 - r)) and its
# worst is r^(m - 1). Rates are illustrative.
print("cold-day loss c   cycle growth g=0.3, w=10   break-even g*")
for c in (0.01, 0.03, 0.05, 0.10):
    print("%14.2f   %25.3f   %13.3f"
          % (c, 1.3 * (1 - c) ** 10, (1 - c) ** -10 - 1))
R = 0.97
print("restate every m layers at r = %.2f:" % R)
print("    m   mean retention   worst")
for m in (1, 2, 4, 11):
    mean = 1.0 if m == 1 else (1 - R ** m) / (m * (1 - R))
    print("%5d   %14.3f   %5.3f" % (m, mean, R ** (m - 1)))

Single source and citation

Fan, Y. (2026). Mid-Context Forgetting in Long-Context Attention: Positional Bias and the Threshold for Lossless Segment-Summary Relays (Version 1.0.0), Chapter 7: The Classics Examined: Mencius on One Day of Sun and Ten Days of Cold, and Ebbinghaus on Memory. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-255 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026255,
  author    = {Fan, Yafu},
  title     = {Mid-Context Forgetting in Long-Context Attention: Positional Bias and the Threshold for Lossless Segment-Summary Relays},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book