There is a small experiment from the Gegenfurtner lab in Giessen that does an unusual thing. They show a viewer a photograph of a banana on a uniform gray field, then ask the viewer to adjust the banana until it appears perfectly achromatic. The setting the viewer chooses is not gray. It is faintly bluish, because the viewer is unconsciously cancelling out a yellow tint that the banana never had. The image was already gray. The yellow came from inside the head of the person looking at it.
That stored, object-specific colour is what perception researchers call a memory colour. Bananas are yellow, grass is green, oranges are orange, skin tones cluster within a narrow band. Your visual system has logged those statistical regularities over a lifetime and now reaches for them whenever it recognises one of the objects, biasing what you actually see toward what you remember things looking like. The same machinery quietly bends every guess you make in a colour-memory test, including the ones we run on this site.
Where the term comes from
The idea is old. Ewald Hering described it informally in 1920, observing that a familiar object viewed through tinted glass looked closer to its remembered colour than to its physically present one. For most of the twentieth century the effect was treated as a curiosity of high-level cognition, separate from early colour perception. The modern revival is usually traced to Hansen, Olkkonen, Walter and Gegenfurtner’s 2006 paper Memory modulates color appearance, which ran the gray banana procedure with photographic stimuli and a calibrated display, and showed the bias survives careful controls (Hansen et al., 2006). A randomly-shaped blob in the same yellow as the banana did not produce the same shift. The object identity was doing real perceptual work.
What the brain is actually doing
The most useful way to think about memory colours is as a prediction. The visual system never has enough information from the retina alone to recover the “true” colour of an object, because the light reaching the eye is a product of the surface reflectance and the illumination, and you only ever receive the product. A banana under tungsten light, a banana under noon daylight, and a banana under shade all send very different signals down the optic nerve. Yet a banana still looks yellow in all three.
The trick is colour constancy: the brain estimates the illumination, divides it back out, and recovers something close to the underlying reflectance. Memory colours are the part of that machinery that uses what you know about the object. If the visual system has already decided “banana”, it has access to a strong prior on what bananas reflect, and it can fold that prior into the estimate. For familiar, canonical objects this prior is so well-trained that it can shove the perceived colour away from the physical signal by several units of CIE Lab. Witzel and colleagues measured the size of the shift across nine familiar objects and found the bias was strongest for fruits and skin and weakest for objects whose colour varies in the real world, such as cars or cloth (Witzel et al., 2011).
Why this matters for a color memory game
The rounds on the Daily challenge and Solo never show you a banana. The targets are unlabelled colour swatches drawn from across the spectrum, chosen precisely because they sit between named categories. So the classical memory-colour effect, where a recognised object biases its own perceived colour, does not apply directly.
What does apply is the same underlying mechanism in a more general form. Bae and colleagues (Bae et al., 2015) showed that when a colour has to be held in working memory for a few seconds and then reproduced, the recalled value drifts systematically toward the centre of the nearest colour category. A colour halfway between teal and blue gets pulled toward prototypical teal or prototypical blue, depending on which label the viewer attached during the brief encoding window. The category boundary is not arbitrary. It tracks the boundaries of the basic colour terms in the viewer’s language, and the drift is larger after longer delays. This is memory colour working on the level of named hues rather than named objects, and it is the dominant source of error on a four-second reproduction task.
On the scoring side the bias has a sharp shape. A target that sits squarely in the middle of one of the eleven basic colour terms tends to be recalled accurately, because the gist label and the actual hue agree. A target that sits on a category boundary gets recalled badly, because two different labels are equally plausible and the brain picks one and then drifts toward the centre of it. If you have ever felt that the colour you saw “has to be” teal but you cannot decide between teal and aqua, you have already experienced the effect. Whichever label wins, your slider settles closer to that label’s prototype than to the actual target.
How memory colours help and hurt at the same time
The reason this machinery exists is that it makes perception more useful in the real world. A grocer who sees the banana in the back of a cool-lit shop as the same yellow it would be on a sunlit cart can grade it consistently across lighting conditions. A doctor who looks at a patient’s skin under fluorescent lighting still reads the same flushed pink that would appear at the bedside in daylight. Colour constancy and the memory-colour bias that supports it are why colour vision is informative about objects, not just about light.
Inside a reproduction game the same trick is a tax. You are being asked to recover the physical signal, but the visual system has already done its category-based smoothing on the way in. You can feel the smoothing happening. When the slider screen first appears, the colour in your head feels like a clear, nameable family member of some basic colour. After a couple of seconds of trying to dial it back, the original signal is already gone and only the family label is left to work with. Whatever you produce on the slider is closer to the prototypical colour of that family than to the actual target.
How to fight your own memory colours during a round
There are three practical moves, each of which exploits a different part of the mechanism. None of them defeat the effect, but each one shrinks it.
Pick a non-basic label during the flash. If you encode the colour as “teal” or “green”, you will recall it as the centre of teal or the centre of green. If you encode it as “teal but darker than usual, leaning blue”, you give the recall stage three pieces of information instead of one and the drift toward the prototype is partially cancelled. A richer colour vocabulary is the single biggest predictor of accuracy on the reproduction format, and it is the only part of the task that genuinely trains with practice. We wrote about how to build that vocabulary in Train your eye for color.
Commit to a saturation and brightness, not just a hue. The category bias is much stronger for hue than for the other two slider axes, because basic colour terms are organised around hue first. If you encode “medium dark, muted” alongside the hue label, you protect the two sliders that the bias does not cover, and you can usually recover most of the round even when the hue itself drifts.
Move the sliders as soon as the screen appears. Visual working memory decays within seconds, and the longer you sit on the slider screen unsure of where to start, the more the gist label takes over from the original colour. The first position you put the sliders in, before you start second-guessing, is usually closer to the truth than the position you settle on after thirty seconds of fiddling. We score with CIEDE2000, which is perceptually uniform, so a confident wrong move is rarely much worse than a hesitant slightly-less-wrong one, and the speed bonus on most variants tips the balance further toward acting early.
Where this fits in the wider science
Memory colours sit at the intersection of two long-running debates in colour science. The first is whether colour perception is bottom-up (driven by retinal signals) or top-down (driven by prior expectations). The Hansen banana experiment is the cleanest case for “both”: the same physical stimulus appears different depending on what you believe the object is. The second debate is whether colour categories are universal or learned. The category-drift effect Bae and colleagues measured tracks the colour terms in the viewer’s language, and similar studies in Russian (which has two basic terms for blue) and Greek find the boundaries shift to match (Winawer et al., 2007). So your memory colour for “teal” is partly a record of the wavelengths that your ancestors found worth naming, filtered through what your specific language picked out.
That is also why a color memory test is not a pure measurement of either perception or memory. It is a measurement of how cleanly you can pass a colour through your own labelling system without losing too much of the between-category detail. The score is high when your vocabulary is precise enough and your labelling is fast enough that the category drift has nothing soft to grab onto. The score is low when the only label you had time to attach during the flash was a coarse one and the drift has had four seconds to do its work.