B = −0.5 × ln(16) = −0.5 × 2.773 ≈ −1.386.
A negative activation isn’t a special failure state — it’s just a number on a continuous scale that feeds into a retrieval-probability formula (a logistic function of B, in the full architecture): the more negative, the slower and less reliable the retrieval, down to outright failure. Compare a chunk used 1 day ago instead: B = −0.5 × ln(1) = −0.5 × 0 = 0, clearly stronger. The Grünfeld’s name, at 16 days cold, sits well below that — retrievable, but sluggish, exactly the “it’s on the tip of my tongue” feeling.
Why decay is rational, not a design flaw. This is ACT-R’s most provocative empirical claim, from Anderson & Schooler’s 1991 rational analysis: memory’s actual forgetting curve (fast initial decay, long slow tail — a power law) matches, almost exactly, the statistics of when information actually becomes needed again in the real world. Words that appeared in a newspaper headline yesterday are far more likely to reappear today than a word from a headline a year ago; email you received an hour ago is far likelier to be relevant right now than email from three years back. A memory system that keeps everything equally retrievable forever is not obviously better — it pays a retrieval-time and interference cost on every single lookup, searching a haystack that’s mostly hay you’ll never need again. Decaying activation is a bet, tuned by evolution and daily use, that recency and frequency of past use predict future need — and empirically, for most of what a mind encounters, that bet pays off. Forgetting the Grünfeld’s name isn’t memory malfunctioning; it’s memory correctly downgrading something increasingly unlikely to matter today, freeing retrieval effort for what’s more likely to.
Spreading activation is the correction for when that bet would otherwise fail: base-level decay alone would leave the Grünfeld equally hard to retrieve regardless of context — but the moment the player sits down for a chess game (goal context now includes “chess,” “openings”), spreading activation from those active chunks boosts every chess-adjacent chunk, the Grünfeld included, faster than base-level decay alone would predict. The two mechanisms together explain something base rate alone can’t: why a stale memory can suddenly feel instantly available the moment the right context shows up — cue-dependent retrieval, not just time-dependent decay.
Where this goes: ACT-R modeled the mind’s own memory dynamics as an architecture. SOAR, its great rival cognitive architecture, was built from a different first commitment — that all cognition is problem-space search — and reaches surprisingly similar territory from the opposite direction. That comparison starts next stage.