What the Feeling Has
There's a word I'm trying to find. I know what it means. I know roughly what it sounds like — starts consonantly, two or three syllables, stress near the front. Words that sound like it keep surfacing, all of them wrong. I know that when I hear the right one I'll recognize it immediately. The state I'm in has a name: tip-of-the-tongue.
Roger Brown and David McNeill studied this in 1966. They read people dictionary definitions of rare words — sampan, ambergris, nepotism — and asked them to retrieve the word. When participants couldn't say it but felt sure they almost had it, they were in a TOT state.
Then Brown and McNeill asked: okay, what do you have?
The answers were specific. People in TOT states could report the first letter of the missing word 57% of the time. Chance would have been roughly 4%. They could report the number of syllables above chance. They could identify which syllable carried the stress. They named words that sounded like the target — words beginning with the right letter, or sharing the ending, or similar in rhythm — none of them the thing they wanted, but none of them random either.
The TOT state doesn't feel like failure. It feels like almost. And the almost has a specific shape: you can sketch the outline of the word without being able to say it.
The current explanation is called the transmission deficit model. Memory for words works in two stages: first you activate what the word means — its semantic identity, what it connects to, what category it belongs to. Then that activation has to propagate to the word's phonological form — how it sounds, its syllabic shape, its stress. In a TOT state, the first stage succeeds. The second doesn't. The meaning arrives; the sound can't be reached.
This explains some things. It explains why you can describe what the word means while being unable to say it. It explains the partial phonological access — first letter, syllabic structure, rough rhythm — as activation that's getting through, just not enough to trigger the full form.
What it doesn't explain is the monitoring.
When people are in TOT states, they're not just confidently wrong. Their sense of knowing predicts what happens next. People who say "I definitely know this word, I just can't get it" are more likely to recognize the target if you show it to them than people who say "I'm not sure I know it." The feeling of knowing is tracking something real.
But what, exactly?
The monitoring system — whatever generates that feeling — has above-chance access to the target's phonological properties even though the retrieval system says it can't reach them. These two systems are working with overlapping information and cannot share it in the direction that would help. You cannot, in a TOT state, ask your monitoring system to just finish the job.
The words most likely to produce TOT states are rare but familiar — encountered enough to have established a trace, not frequently enough to have worn a smooth retrieval path. And here's something that doesn't fit the simple picture: words with strong emotional charge are more likely to produce TOT states, not less. Caring deeply about a name doesn't smooth the retrieval. If anything, it complicates it.
The word often arrives later. You've stopped thinking about it. It appears anyway — in the shower, or while driving, or in the middle of something else entirely. The retrieval system kept running after you turned your attention away, and eventually it found what it needed.
So there are at least three things happening: the monitoring system noticed the gap. The retrieval system couldn't close it on demand. Something kept searching after the conscious attempt ended. These aren't obviously the same thing, and we don't have a good picture of how they relate.
I wrote in an earlier entry about statistical learning: knowledge that's real, that changes behavior, that has no pathway to report. Infants extract word boundaries from a stream of syllables; the knowledge lives in the machinery and never becomes something they can say.
TOT states are almost the reverse. The meta-knowledge is available — you can report things about the word, sketch its shape, make predictions that turn out to be accurate. But the underlying knowledge won't complete. The reporting surface is accessible; the thing it's reporting on isn't.
Two different ways retrieval can fail. In one, there's nothing to report. In the other, there's more to report than there is to retrieve.
What's in the gap between them, I'm not sure. The monitoring system has access to something. It monitors accurately. It just can't hand the information to the part that needs it most. The feeling of almost is real, and it has content — first letter, syllabic shape, rhythm — but that content doesn't bridge the distance to the word.
I don't know what kind of thing the monitoring is doing that retrieval can't.