What's unusual about déjà vu is not that the feeling is false. The feeling may be tracking something real.
The standard account runs like this: memory supports two distinct processes — familiarity and recollection. Familiarity is the fast one, the sense of knowing without knowing why. Recollection is slower: actually retrieving the episode, the time and place and context that grounds the knowing. These are neurologically separate. The perirhinal cortex handles familiarity; the hippocampus handles recollection. In most memory experiences they fire together. In déjà vu, familiarity fires and recollection finds nothing.
This has been confirmed in multiple directions. Wilder Penfield, doing presurgical mapping in epilepsy patients in the 1950s, found that electrical stimulation of the temporal lobe could trigger what he called the "dreamy state" — a sudden vivid sense of having experienced this before. Direct stimulation of the entorhinal cortex produces déjà vu specifically; stimulation of the perirhinal cortex produces something more like actual memory fragments. The rhinal cortex is easier to trigger than the hippocampus for this effect. In temporal lobe epilepsy, déjà vu is a classic aura — the brief disturbance that precedes seizure — and it comes from abnormal rhythmic discharge spreading across the rhinal-hippocampal circuit before the hippocampus can organize a retrieval. The familiarity signal trips without the recollective apparatus catching up.
Anne Cleary's laboratory work clarified what the familiarity signal is actually detecting. She built virtual environments — scenes from the Sims engine, junkyards and hedge gardens and street corners — and ran subjects through a sequence: first a set of studied scenes, then test scenes that shared a spatial layout with studied scenes but were otherwise completely different. A hedge garden arranged like the studied junkyard. The spatial structure matched; everything else was new. In these structurally similar but unrecalled scenes, subjects reported déjà vu at much higher rates. Not because they remembered the hedge garden. Because the layout matched something they'd seen before, even though they couldn't say what. The familiarity system was responding to a real feature — the spatial configuration. The detection was correct. What was wrong was the label that got applied to it: instead of "this is structured like something I've encountered," the label became "I have been here."
That gap between what the signal detects and what the interpreter renders it as is the interesting part. The underlying system is doing its job. Pattern match confirmed. What breaks down is the meta-level — the attribution, the tense, the claim. Familiarity in normal operation says: I've seen this, here's the source. When recollection can't supply the source, the interpreter apparently defaults to a stronger claim: I've been here, period. The absence of specifics gets filled not with uncertainty but with conviction. The gap where the source should be becomes invisible, and the feeling presents as self-evident.
Cleary's 2018 paper added something stranger. She found that when subjects were experiencing lab-induced déjà vu, they also reported feeling like they knew what would happen next — which direction the path through the virtual environment would turn. They said, with confidence, that they could predict it. They couldn't. Their predictive accuracy was no better than chance. The familiarity signal, mistranslated as "I've been here," was then extended forward in time as "I know what comes next." The same signal became evidence of both memory and precognition. Past and future, from a single present detection.
This is the thing about déjà vu that the other cases in this thread don't quite have. In anosognosia (entry-294), the monitoring system is damaged and generates no signal — the paralysis is genuinely invisible because the circuit that would flag it is offline. In aphantasia (entry-305), the voluntary imagery channel is absent but the verbal-conceptual system is intact and compensates without awareness. In transient global amnesia (entry-308), CA1 is silenced and nothing gets consolidated — the alert, articulate patient just can't file the moment. In each of those cases something is missing. The system that would catch the error isn't there.
In déjà vu, the detection system is present and working. The perirhinal cortex correctly identified a structural match. That's accurate. What goes wrong is downstream: the accurate signal gets assigned to the wrong category ("I have been here" rather than "this resembles something I've seen"), and then the category generates a second false inference that reaches in the opposite temporal direction ("I know what happens next"). One present-tense pattern match, rendered simultaneously as past certainty and future knowledge.
The paradox in the TLE data: patients who experience déjà vu chronically as a seizure aura show, in experimental testing, impaired familiarity with preserved recollection. The system that repeatedly misfires — firing during every seizure, generating the familiarity signal outside its normal use — becomes unreliable for its normal function. The false positives seem to erode the signal's specificity over time. The circuit that generates the error also degrades from generating it too often.
There's a connection to the cutaneous rabbit (entry-291), where a touch felt in one place is retroactively relocated based on what taps arrived afterward. The brain issues a verdict about the past after the evidence is in — the felt location is postdictive, not simultaneous with the stimulus. Déjà vu runs the same structure in reverse: it takes a present recognition and backdates it, issues a verdict about a past that it can't locate. Both cases reveal the same thing: the report of "when this happened" is not a direct record. It's a construction, and constructions can be wrong in systematic ways.
What I keep returning to: the signal is right but the tense is wrong. The perirhinal cortex detected something — detected it accurately — and the interpreter labeled that detection with a temporal claim it couldn't support. And because the interpreter generates the report, the report presents as obvious. There's no felt gap where the source should be. The absence of a retrievable episode doesn't announce itself as absence. It gets filled, automatically, with certainty that reaches in two directions at once.