loading...
Accuracy: rolling (last 24 trials) and cumulative, against the always-no baseline
running...
The science
Working memory holds a few items for seconds while you use them. Miller (1956) put the limit near seven items; Cowan (2001) argued for about four. The N-back task (Kirchner 1958) asks whether the current letter equals the one N trials ago.
The model (identical to the 3D page: same file, same constants, same random draws):
K slots; a resident trace has strength s, ds/dt = -d s (s = exp(-d t)); a new letter enters with s = 1
full store: the slot with the smallest s is replaced
true match: P(say match) = sigmoid(6 (tau - c - 0.12))
no match: P(say match) = sigmoid(6 (c - 0.55))
tau = trace of the target letter, c = mean strength of the other traces
The decision is a signal-detection rule (Green and Swets, 1966); the hit and false-alarm rates give d' = z(H) - z(F) with a log-linear correction. The task forces a true match in 38 % of trials, so a participant who always answers "no" scores about 1 - 0.4055 = 59 % - the orange baseline.
Why capacity and decay matter. A 2-back target is two trials old, so its trace is exp(-d x 1.1 x 2); with d = 1 that is 0.11 and the rule almost never accepts it. Slower decay or a bigger store keeps the target strong enough.
Checks in the page: decay vs exp(-d t) (two frame sizes), the true-match fraction vs its closed form, hit and false-alarm probabilities vs the sigmoid, the live engine vs an independent re-implementation with a different random generator (300,000 trials), and a hand-computed d'.
Honest simplifications. Only the slot idea is modelled (no phonological loop, rehearsal, chunking or central executive); exponential decay plus eviction is deliberate - whether forgetting is decay or interference is debated; all constants (6, 0.12, 0.55, 1.1 s trial interval) are illustrative, not fitted to human data; eviction ties go to slot 0. The tiles, glass slots and brain are Blender 5.2 renders and illustrative only. Citations are quoted from memory: spot-check.