loading...
Accuracy: rolling (last 24 trials) and cumulative, against the always-no baseline
Experiment result (A: accuracy against capacity K for 1-, 2- and 3-back; B: accuracy, hit and false-alarm rate against decay)
running...
How to use
A letter appears every trial. The simulated participant answers MATCH if the letter equals the one shown N trials earlier (the gold frame on the belt) and NO MATCH otherwise. Every letter is encoded into a glass slot; the tube beside each slot is its trace strength, which decays exponentially, and when the store is full the weakest trace is thrown out. Orbit with the mouse, zoom with the wheel; the camera buttons fly to the stream, the store, the screen, the brain and the response box, and Tour cycles through them. Pick a preset or change K, N, decay and speed; Step runs one trial, Space pauses, R resets, S steps. The thin amber arc is the retrieval beam from the target tile to the slot that still holds its trace (no arc: the target was already forgotten). The sparks show a new letter being encoded. The model does not learn; accuracy only reflects the store.
The science
Working memory is the small amount of information you keep active for a few seconds while you use it. Miller (1956) put the limit at about seven items; Cowan (2001) argued that the real limit of the focus of attention is about four chunks. Baddeley and Hitch (1974) describe working memory as several parts (a phonological loop, a visuospatial sketchpad, a central executive, later an episodic buffer); items in the phonological store fade within about two seconds unless rehearsed.
The model. This page runs the same slot model as the 2D lab (constants, random draws and trial order identical):
store: K slots; a resident trace has strength s, ds/dt = -d s (so s = exp(-d t)); a new letter enters with s = 1
full store: the slot with the smallest s is replaced
target trace τ = s of the slot holding the letter shown N trials ago (0 if it was forgotten)
confusion c = mean s of the other resident traces
true match: P(answer MATCH) = σ(6 (τ - c - 0.12)) no match: P(MATCH) = σ(6 (c - 0.55))
σ(x) = 1 / (1 + e-x), one trial every 1.1 time units (the speed slider scales time)
This is a signal-detection decision (Green and Swets, 1966): the hit rate H and the false-alarm rate F give the sensitivity d' = z(H) - z(F), with a log-linear correction for 0 or 1. The task generator forces a true match in 38% of the trials (so the true-match fraction is 0.38 + 0.62 x 0.0411 = 40.5%); an observer who always answers NO is therefore right 59.5% of the time, which is the orange baseline in the accuracy graph.
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 only 0.11 and the decision rule almost never accepts it; slower decay (Experiment B) or a bigger store (Experiment A) keep the target resident and raise the hit rate, while a store full of strong, similar traces raises the false alarms (confusion). The tube gauges and the graph of strength against age show the decay directly.
Numerical checks run in the page: (1)-(2) a decaying trace matches exp(-d t) exactly, for 1/60 and for 0.55 frames (frame-rate independence); (3) the fraction of true matches in 60,000 engine trials agrees with the closed form 0.4055; (4)-(5) with a pre-loaded store the hit and false-alarm frequencies agree with the sigmoid formulas (40,000 repeats each, binomial error about 0.2%); (6) the accuracy of the live engine at the current settings agrees with an independent re-implementation (different code structure and random generator, 300,000 trials), typically within half a percent.
Honest simplifications. (a) Only the slot (Cowan-style) idea is modelled: no phonological loop, no rehearsal, no chunking, no central executive. (b) Exponential decay plus eviction is a deliberate simplification; whether forgetting is decay or interference is debated, and this model mixes both in a crude way (the confusion term is interference). (c) The constants 6, 0.12 and 0.55 and the 1.1 time-unit trial interval are illustrative, not fitted to people; the time unit is not calibrated to seconds. (d) Eight letters only; trace values are exact and noise-free, apart from the random choice of letters and answers. (e) The reward, loss and entropy values from the 2D page are telemetry of the run, nothing is learned. (f) The brain is a stylised object: working memory engages a frontal-parietal network and the glow only shows the store load; it is not an image of real data. (g) Ties in eviction go to the lowest slot index, as in the 2D page.
Sources to spot-check (quoted from memory, flagged): Miller 1956 (Psychol Rev 63); Kirchner 1958 (J Exp Psychol 55) for the N-back task; Baddeley and Hitch 1974; Cowan 2001 (Behav Brain Sci 24); Green and Swets 1966; Oberauer 2002 (J Exp Psychol LMC 28). Educational model, not a diagnostic or training tool.