Color and Working Memory: Why the Shade Slips Away
What happens to a color in the seconds after it disappears, and why the error is not random.
Short answer: holding a color in mind for even a couple of seconds costs you precision, and the loss is not simply blur. Working memory keeps only a small number of colors with usable fidelity, the fidelity depends on which color it is, and the reported color drifts in predictable directions: towards the prototype of its category, and towards the average of whatever else you have been seeing. That is why reproducing a color is a much harder job than comparing two colors side by side.
The task that made this measurable
Most of what is known here comes from one deceptively simple experiment: delayed estimation. A colored patch appears briefly, then vanishes. After a short delay the observer picks the remembered color from a continuous wheel. Because the response is continuous rather than right or wrong, the whole distribution of errors is available: how wide it is, whether it is centred on the target, and whether the misses fall on one side.
That distribution turns out to be far more informative than an accuracy score. A person can have a small average error and a large systematic bias, or the reverse, and those two situations call for completely different advice. This is the core reason we report reproduction games differently from discrimination games.
Capacity: a small number, honestly argued about
An influential analysis of delayed estimation concluded that observers hold a limited number of items in visual working memory, on the order of three, each with a resolution that does not keep degrading as more items are added. Beyond the limit, the model says, you are not remembering everything badly, you are remembering a few things and guessing the rest. The signature is a response distribution that is a mixture of a tight peak around the target and a flat carpet of random responses.
That interpretation, usually described as a slots account, has been contested ever since by accounts in which a continuous resource is shared out across items. We are not going to settle it here, and you should be wary of any site that claims it has. What survives the argument is the practical part: performance falls off sharply after a small number of colors, and the fall is steep enough to feel. If you are holding four colors you are already past the comfortable range, which is roughly where Huemory starts to bite.
Precision is not a single number about you
It is tempting to think of color memory precision as a personal constant, the way people think of eyesight as a number on a chart. Delayed estimation says otherwise. When researchers looked at the response distributions for different target colors, the widths differed substantially from color to color, and, tellingly, the pattern was correlated across observers: the colors that were hard for one person were hard for the others. Further experiments showed that stimulus-specific responses persist under higher memory loads, and that at least part of the specificity arises in perception before being carried into memory.
The consequence for anyone tracking their own scores is direct. An average error across a mixed set of colors is a blend of your ability and the particular colors you happened to draw. If today's targets sat in a region that is intrinsically noisy, your number will look worse for reasons that have nothing to do with you. This is why our guide to recording and evaluating your scores insists on trends over a fortnight rather than day-to-day comparisons.
The biases: where the memory goes when it moves
Two systematic pulls dominate, and they can operate at once.
| Bias | What it does | What makes it stronger |
|---|---|---|
| Categorical pull | The remembered color drifts towards a prototypical example of its color category, a typical blue or a typical green, rather than staying at the exact shade. | Longer delays; colors near a category boundary; having to name the color. |
| Central tendency | Reports drift towards the mean of the set of colors you have been seeing recently, so extremes are pulled inward. | Noisier stimuli, longer delays, anything that weakens the memory relative to prior expectations. |
| Capacity overflow | Beyond a small number of items, some colors are effectively not stored at all and the response is close to a guess. | More items to hold; distraction during the delay. |
The categorical pull has its own page here, color memory and recall, because it is the one people meet first when they try to describe a color to a shop assistant. The central tendency bias is less famous and just as important: work manipulating internal and external noise showed that as the memory signal weakens, estimates lean harder on the average of the stimulus distribution. In plain terms, when your memory is uncertain your brain falls back on what is typical, and it does so without telling you.
A model that combines a fine-grained memory of the particular shade with a coarse memory of its category accounts for the pattern better than either component alone, and it also explains why some colors appear more memorable than others: colors that sit near a category centre get help from the category, while colors near a boundary get pushed around by it.
What this means for the games
Hueecho is delayed estimation with a friendlier interface. A color appears, it goes, and after a pause you rebuild it. Expect a bias, not just noise, and expect the bias to grow with the delay. If your reproductions land consistently on one side of the target, that is a finding about you rather than a bad day.
Huemory adds load: a sequence rather than a single color. This is where the capacity limit shows up as a cliff rather than a slope. Most people find that their accuracy is comfortable up to a small number of items and then collapses, and that the collapse point is fairly stable across sessions.
Huespan asks for something subtly different: reproducing an interval rather than a single color. That means holding a relation in mind, and relations appear to be stored differently from absolute values. Our page on color midpoint and interval perception covers what makes that its own skill.
Across all three, the comparison that matters is with Huefork, where the two colors are present together. The gap between your discrimination threshold and your reproduction error is a rough measure of what memory is costing you, and it is usually larger than people expect.
Practical ways to lose less
Verbalise, but carefully. Saying "medium-saturation blue-green, quite light" gives you a second, categorical handle on the color. It also strengthens exactly the categorical pull described above, so it helps most when the color is near a category centre and hurts when it sits on a boundary.
Anchor to a dimension. Rather than trying to store the color as a whole, note where it sits on the dimension you find easiest, usually lightness, and reconstruct outward from there. The hue, saturation and lightness vocabulary is what makes this possible.
Do not fill the delay. Looking at something else during the gap, especially something colored, costs you more than the delay alone.
Keep the display constant. A shifted white point between presentation and response introduces an error that has nothing to do with memory, since your adaptation state has moved. That is a color constancy problem wearing a memory costume.
Read your bias before your average. A systematic offset can be corrected deliberately. Random spread cannot. Look at the direction of your misses first.
Watch your own memory bias appear. Hueecho shows a color, takes it away, and asks you to rebuild it after a pause. Play it a dozen times and look at which side of the target your answers land on. That direction is the interesting part.
Frequently asked questions
Why can I match a color perfectly but not remember it?
Because matching and remembering use different resources. In direct comparison you can exploit the shared edge between two patches. Once the color is gone you are comparing against a stored representation that is limited in number, limited in precision, and biased towards the category the color belongs to.
How many colors can I hold at once?
Fewer than most people expect. Work modelling delayed estimation concluded that observers store a small number of items, on the order of three, and that beyond that limit they guess rather than remember all of them imprecisely. Whether memory really works as discrete slots is still argued about, but the small number is not.
Does the delay matter?
Yes, and not only by adding noise. Longer delays and noisier stimuli increase the pull of systematic biases, including a central tendency bias where reports drift towards the average of the colors you have been seeing. Waiting does not just blur the memory, it moves it.
Are some colors easier to remember than others?
Clearly yes. Delayed estimation studies found that different target colors produce response distributions of very different width, and that this pattern is consistent across observers. Part of that specificity arises in perception and is then carried into working memory.
Is my memory error random?
No, and that is the useful part. Random error shrinks with more trials; systematic bias does not. If your reproductions consistently land on the same side of the target, that is a bias you can learn to correct for, which is worth more than a slightly lower error average.
Why does Hueecho feel so much harder than Huekle?
Because Huekle lets you compare against a target that is still there, while Hueecho takes the color away and inserts a delay. The task changes from discrimination to reproduction, and reproduction inherits every limit of working memory on top of your perceptual precision.
Does holding more than one color make each one worse?
That is exactly the question the field has been arguing about. The evidence is clear that performance degrades as the number of remembered items rises; the disagreement is about whether precision is shared out continuously or whether a small set of items is stored well and the rest not at all.
Can I train color working memory?
You can certainly get better at the task, and you can learn to recognise and correct your own biases, which helps immediately. Whether the underlying capacity changes is a much bigger claim, and we do not make it. Our page on training color vision explains why we are cautious.
Keep reading
- Color memory and recall: the categorical pull, in plain language.
- The just-noticeable color difference: the perceptual floor that memory sits on top of.
- Color midpoint and interval perception: why remembering a relation is its own task.
Sources: Bae, G.-Y., Olkkonen, M., Allred, S. R., Wilson, C. and Flombaum, J. I. (2014), Stimulus-specific variability in color working memory with delayed estimation, Journal of Vision 14(4), 7; Zhang, W. and Luck, S. J. (2008), Discrete fixed-resolution representations in visual working memory, Nature 453(7192), 233; Bae, G.-Y., Olkkonen, M., Allred, S. R. and Flombaum, J. I. (2015), Why some colors appear more memorable than others: a model combining categories and particulars in color working memory, Journal of Experimental Psychology: General 144(4), 744; Olkkonen, M., McCarthy, P. F. and Allred, S. R. (2014), The central tendency bias in color perception: effects of internal and external noise, Journal of Vision 14(11), 5.
Understand it, now train it. The daily games stay free forever. A Personal Pass adds progress tracking (trends + export); Teacher / School tools let you assign it and follow a whole class. For teachers & self-learners →