Solution: Why 'Doctor' Makes You Faster to Recognize 'Nurse'

Numeric answer: A(nurse) = 1.75

A(nurse) = B(nurse) + Σⱼ(Wⱼ × Sⱼ,nurse)
         = 0 + [0.5 × 2.0] + [0.5 × 1.5]
         = 0 + 1.0 + 0.75
         = 1.75

Nurse has never been directly retrieved recently (B = 0), yet it carries substantial activation purely from spreading — because two chunks strongly associated with it (“doctor,” “medicine”) happen to be occupying working memory right now.

Part 2 — Bread’s activation: 0.1, and the comparison

A(bread) = 0 + [0.5 × 0.1] + [0.5 × 0.1] = 0 + 0.05 + 0.05 = 0.1

Nurse (1.75) vs. bread (0.1): a gap of 1.65 in activation, purely a function of which chunks happen to be semantically associated with what’s currently active in working memory — nothing about either word’s own recency or frequency of use differs in this scenario (both have B = 0). Since Lesson 10 established that higher activation predicts faster, more reliable retrieval, this gap directly predicts nurse being recognized substantially faster than bread right now — which is exactly the empirical priming result Meyer & Schvaneveldt found: “doctor” primes “nurse” via spreading activation along a real associative link; “doctor” does not meaningfully prime “bread” because no comparable link exists (or it’s weak, as modeled by S = 0.1 here).

Part 3 — Why priming decays

If spreading activation permanently strengthened every associate of everything you’d ever attended to, activation would only ever accumulate — every chunk in memory would slowly ratchet toward maximal activation as it picked up residual boosts from an ever-growing history of things once associated with it, however distantly or briefly. Two things would break:

1. Loss of discrimination. Activation is only informative if it distinguishes “relevant right now” from “not relevant right now.” If old associations never faded, current context would be swamped by irrelevant residue from every prior context, and retrieval would stop being fast or selective — everything would eventually look equally activated.
2. No signal about current relevance. The whole point of the Wⱼ term is that it reflects what’s occupying attention now — as soon as doctor/medicine leave working memory (attention moves elsewhere), their Wⱼ should drop toward zero, because they’re no longer evidence about what’s currently relevant. A system where past attention never releases its grip on activation couldn’t track a changing world or a changing task.

Priming’s fast decay (seconds to low minutes, in the human data) is exactly tuned to track “what’s relevant to the current context,” not “what have I ever thought about” — that second, durable kind of strengthening is what base-level activation (B, from Lesson 10’s repeated-use decay curve) is for instead. The two terms of the activation equation divide the labor: B handles slow, durable learning from repeated use; the spreading term handles fast, transient relevance to the current moment.

Part 4 — A candidate analogue in LLM agent harnesses

The closest structural analogue is the attention mechanism inside a transformer itself (the “attention” the fp and ml tracks describe as a parallel soft lookup): when processing a token, the model computes attention weights over all other tokens in context, and tokens strongly related to the current one (by learned associative patterns, not an explicit symbolic link table) get weighted more heavily in producing the next representation — a direct computational parallel to Wⱼ × Sⱼᵢ summed over active sources. The similarity: both are automatic, “for free” mechanisms that boost related information without an explicit, separate retrieval step being invoked.

The important difference: transformer attention operates within a single forward pass over the current context window — it doesn’t persist or decay across turns/interactions the way human priming does across seconds to minutes; each new prompt essentially recomputes attention from scratch over whatever’s in the context window at that moment, with no residual “priming” carried over unless the harness explicitly re-includes prior content. A harness that wanted a genuine priming-like effect across separate calls would need to build that persistence deliberately — e.g. a scratchpad or memory store that re-surfaces recently-relevant items into future contexts — since the base transformer mechanism, unlike human spreading activation, doesn’t carry activation forward on its own.

The pattern

Rule: activation isn’t a single number describing “how strong is this memory” in isolation — it has a durable component (repeated use) and a transient component (what’s currently active and associated). Priming is the transient component made visible: chunks connected to what’s in mind right now get temporarily easier to retrieve, and that boost is supposed to fade once context moves on.

Where this goes: next lesson turns from associative memory to a different cognitive bottleneck — attention as selection: given that only so much can be “in working memory” (and therefore able to spread activation) at once, how does the mind decide what gets in?