George Miller, cognitive psychologist at Princeton University, published one of the most-cited papers in psychology in 1956. He compiled experiments from various laboratories and found a striking pattern: whether tones, digits, letters or syllables — memory fails consistently at about seven items. Not five, not fifteen: seven, plus or minus two.
Miller called the number “magical” because it came without explanation. His paper was self-reflectively honest: “I have been persecuted by an integer,” he wrote — haunted by a whole number without knowing why it appears. Half a century of research has changed little. Measurements were refined: Nelson Cowan showed in 2001 that the true capacity without rehearsal is closer to four chunks. But a mechanistic explanation deriving from brain biology why the limit is not twelve or three was absent.
Miller (1956): 7±2 items in short-term memory. This figure includes active rehearsal and is therefore biased upward.
Cowan (2001): Without rehearsal strategies, capacity is ~4 chunks. Four is the “bare” number of working memory.
Shared finding: The limit is real, robust and biological — but its cause was unknown.
Adaptive Holographic Theory (AHT) treats the brain not as a network of firing neurons but as a medium in which wave patterns propagate. Meaning does not arise locally — it is a global field state ψ spreading over the entire connectome graph. Working memory in this framework is not a short-term store; it is a time-limited perturbation δL of the graph Laplacian arising from Hebbian co-activation and decaying at forgetting rate κ:
dψ/dt = −i(L₀ + δL)ψ − γψ + S(t)
d(δL)/dt = −η·F(t)·Re[ψψ†] − κ·δL
N simultaneous working memory contents correspond to N simultaneously maintained attractors in the field ψ. How many can be stable at once? That is the question Miller could not answer — and which can now be posed formally.
The decisive step lies in a question one does not ask intuitively: do attractors compete for energy (|ψ|²) or for amplitude (|ψ|)? The answer follows from biology.
Neurons have a maximum firing rate. The refractory period — the recovery phase after an action potential — is roughly 2 ms, imposing an upper bound of approximately 500 Hz. In the AHT equations this corresponds to the parameter ψ_max: a hard amplitude ceiling enforced by the bistable nonlinearity. Energy could in principle grow without bound; amplitude cannot.
The Hebbian coupling term η·|ψᵢ|·|ψⱼ| is linear in each of the two amplitudes — it strengthens an attractor in proportion to its amplitude, not its energy. Consequently, the shared resource is the total amplitude ‖ψ‖, not the total energy ‖ψ‖².
Step 1 — Equilibrium. At stable co-activation (d(δL)/dt = 0), each of N equally distributed attractors receives Hebbian gain δλₖ = (η·F/κ)·(‖ψ‖/N). The gain is proportional to the allocated amplitude — not to the energy.
Step 2 — Survival threshold. An attractor remains stable only if its Hebbian gain compensates the field damping γ: δλₖ ≥ γ. This yields N ≤ (η·F)/(γ·κ)·‖ψ‖. Note: The proof that the linewidth in the coupled ψ–δL system is set precisely by γ (rather than a more complex joint function of γ, η, and κ) requires a perturbation-theoretic analysis of the coupled system and is an open problem.
Step 3 — Capacity formula. The capacity bound is:
N_max = (η·F) / (γ·κ) · ‖ψ‖_eq
The result rests on a layered estimation procedure. κ = 0.0385 s⁻¹ is calibrated directly from Baddeley’s (1992) measured half-life of 18 seconds — the only parameter with a physical time anchor. η = 0.001 and γ = 0.05 are simulation-internal quantities from experiment P1; their ratio η/γ = 0.02 is dimensionless and physically meaningful (ratio of Hebbian gain to synaptic damping). Translating them into physical seconds requires an explicit biological hypothesis: one simulation step corresponds to one gamma cycle, τ_sim = 1/40 Hz = 25 ms. This assumption is not derived from first principles — it is a testable hypothesis, supported by an STDP consistency check: the empirical STDP time constant τ_STDP ≈ 20 ms gives η = η_phys · τ_sim ≈ 1.25 × 10⁻³, consistent with the P1 value. ‖ψ‖_eq = 12.66 is measured by simulation; an analytic formula for N_active remains a conjecture. With F = 1 (full attention):
N_max = (0.001 × 1) / (0.05 × 0.0385) × 12.66 ≈ 6.6
The result lies within Miller’s empirical range 7±2. The estimate is not read as a precision measurement — with a simulation-measured quantity, a biological hypothesis for τ_sim, and a consistency rather than an independence check for η, a claim to single-percent accuracy would not be honest. What the result shows is that the right order of magnitude follows from the field equations, without post-hoc adjustment.
The most surprising result is not the number 6.6 — it is the double robustness of seven. The capacity formula depends on two independent sets of parameters. The first set — (θ, ψ_max) — is determined by neuron biology: threshold voltage and refractory period. The second set — (η, γ, κ, F) — is determined by metabolic and neuromodulatory constraints: learning rate, synaptic damping, forgetting rate, dopamine level.
Both sets evolved independently, and both converge on N_max ≈ 7. A single mutation in one parameter can shift the number slightly — ADHD, sleep deprivation and stress all act this way — but to fundamentally alter it, both sets would need to shift simultaneously. This explains why seven is stable across cultures, languages, modalities and species.
Calibration I — Neural level: The refractory period (ψ_max) and the bistable threshold (θ) determine how much field amplitude the brain produces at equilibrium. These parameters are bounded by energy consumption and signal speed.
Calibration II — Neuromodulatory level: The Hebbian learning rate (η), synaptic damping (γ) and forgetting rate (κ) determine how efficiently the amplitude budget is converted into stable attractors. These parameters are regulated by dopamine, acetylcholine and noradrenaline.
Emergence: Only the product of both calibrations yields N_max ≈ 7. The number is an emergent consequence, not a hard-coded constant.
Why is capacity persistently lower in some individuals — or reduced under certain conditions? The formula gives direct answers.
In ADHD the parameter F is reduced: the dopaminergic gating function that activates Hebbian learning is weaker. This directly reduces the effective gain η·F, and thus N_max — not because the brain is “worse”, but because the amplitude budget is converted into stable attractors less efficiently. Methylphenidate raises F by blocking dopamine reuptake.
Sleep deprivation increases κ: forgetting rate rises when the consolidation mechanism is impaired. Since κ sits in the denominator, N_max falls. Chronic stress additionally increases γ through persistently elevated cortical excitability.
Ageing primarily impairs prefrontal dopaminergic innervation — both F and η decline slightly. The measured decrease in WM capacity from ~7 to ~5 in later life is consistent with a proportional reduction in the product η·F.
The critical feature common to all these impairments is superlinear sensitivity. The equilibrium amplitude ‖ψ‖_eq does not grow linearly with ψ_max but with an exponent p ≈ 1.3–1.7. Small neuromodulatory adjustments therefore have disproportionate consequences — in both directions.
The formula makes a striking prediction for artificial neural systems. Neuromorphic hardware — chips such as Intel’s Loihi or IBM’s TrueNorth — simulates spiking neurons but has no biological refractory constraint. ψ_max can be set much higher than in the biological neuron.
The naive expectation would be: double ψ_max → double capacity. The simulation experiment shows something different. When ψ_max is doubled, ‖ψ‖_eq does not increase by a factor of 2 but by a factor of 2.5 to 3. The scaling is superlinear: both the stable amplitude per node and the number of active nodes grow simultaneously.
This means that for neuromorphic systems, capacities of 20, 50 or more simultaneous working memory items are physically achievable — provided the ratio η·F/(γ·κ) remains stable. This is not speculation; it is a falsifiable prediction derivable from the theory, testable directly on Loihi hardware.
The result is strong enough for three well-grounded claims. First: the capacity limit of working memory follows from the physical structure of neural field dynamics — it is neither a cognitive convention nor an evolutionary accident. Second: the relevant biological quantity is the maximum firing rate, encoded as an amplitude budget, not an energy budget. Third: the layered estimation — κ from measured data, τ_sim as a biological hypothesis (gamma cycle), η verified by STDP consistency, ‖ψ‖_eq from simulation — yields a capacity of 6.6, consistent with Miller’s empirical range 7±2.
What remains open: the derivation assumes orthogonal attractors and uniform amplitude distribution. Real working memory has overlapping patterns, hierarchical chunking strategies and temporally incoherent signals. A complete treatment of these factors might identify Cowan’s ~4 chunks as the regime without active chunking, and Miller’s ~7 as the regime with phonological rehearsal — both from the same formula, under different assumptions about F(t) and the structure of S(t).
The magical number is no longer magical. It is physical.
Based on AHT v26, experiments measure_field_energy.py and experiment_psimax_capacity.py. Parameters calibrated from P1 experiment, Baddeley (1992) and EEG data. Numerical results reproducible on GPU (NVIDIA RTX 4090, CUDA 12.1).