Lesson 14 established that retrieval strengthens the trace being retrieved. This lesson covers a related but distinct effect: retrieving (or even just perceiving) one item makes related items easier to retrieve too — before you’ve consciously tried to recall them at all.
Terms (standalone):
- Priming: exposure to one stimulus (the prime) speeds up or biases processing of a related stimulus that follows (the target) — even when the prime is irrelevant to the task and the effect is unconscious. The classic demonstration (Meyer & Schvaneveldt, 1971): people judge “nurse” as a real word faster after seeing “doctor” than after seeing an unrelated word like “bread,” in a simple word/non-word judgment task.
- Spreading activation: the proposed mechanism behind priming. Memory is modeled as a network of chunks connected by associative links. When a chunk becomes active (because it’s currently in working memory — you just perceived it or retrieved it), some of its activation “spreads” along its links to associated chunks, raising their activation too — making them faster/easier to retrieve next, even before any deliberate retrieval attempt targets them.
- ACT-R’s activation equation (extending Lesson 10): a chunk
i’s total activation isA_i = B_i + Σⱼ (Wⱼ × Sⱼᵢ)— its own base-level activationB_i(from Lesson 10: how recently and frequently it’s been used), plus a spreading-activation term: a sum, over every source chunkjcurrently in working memory, of that source’s attentional weightWⱼ(how much attention it’s getting — withnsources typically splitting attention asWⱼ = 1/neach) times the associative strengthSⱼᵢbetween sourcejand targeti. Chunks strongly associated with something currently “in mind” get a activation boost — measurably, mechanically, before you’ve retrieved them at all.
Part 1 — Numeric (above): compute the activation
Working memory currently holds 2 chunks — say you just read the words “doctor” and “medicine.”
Each gets an attentional weight Wⱼ = 1/2 = 0.5 (attention split evenly across the 2 sources).
Associative strengths to the target chunk nurse: S(doctor→nurse) = 2.0,
S(medicine→nurse) = 1.5. Nurse’s own base-level activation B = 0 (imagine it hasn’t been
retrieved directly in a while).
Compute nurse’s total activation: A = B + Σ(Wⱼ × Sⱼᵢ).
Part 2 — Compare to an unprimed word
Now compute the activation of an unrelated chunk, bread, under the same working-memory contents
(doctor and medicine both active, W = 0.5 each), with S(doctor→bread) = 0.1,
S(medicine→bread) = 0.1, and the same B = 0.
Given Lesson 10’s rule (higher activation → faster, more likely retrieval, roughly proportional to activation level, with retrieval time decreasing as activation increases), which word — nurse or bread — would you expect to be recognized faster right now, and by roughly how much does the activation gap suggest?
Part 3 — Reveal: why does priming decay, and why does it matter for design?
Priming effects are temporary — the “doctor” boost to “nurse” fades within seconds to minutes as
attention moves on and doctor/medicine leave working memory (their W effectively drops toward 0
as they’re no longer active sources). Propose a reason spreading activation is built to decay
quickly rather than permanently strengthening every associate of everything you’ve ever thought
about. What would go wrong with memory retrieval if primed activation never faded?
Part 4 — Connect to a modern AI system
In a modern LLM agent harness, is there a structural analogue to spreading activation — some mechanism by which processing one piece of information measurably biases or speeds up the processing of related information, without an explicit separate retrieval step? Name one candidate mechanism (in the model’s architecture, or in the harness’s own design) and explain what’s similar and what’s importantly different from the priming/decay story above.