How the radial arm maze separates working from reference memory, what recent studies reveal about Alzheimer's, diet, and aging.
The radial arm maze asks a rodent to remember eight paths at once, without doubling back. In this blog we will look at how the maze works, walk through what recent studies are using it to discover, from neuroinflammation to myelin biology, and look at how automated tracking turns each run through the maze into precise, objective data.
Imagine you have eight errands to run, each down its own street, and each street must be visited once. You set off, and every time you reach the end of a street you have to decide, based on what you remember doing already, which street to try next. Doubling back wastes time; forgetting where you have already been, might mean that the shirt you're shopping for is not available anymore. In a wide range of neurological conditions, from Alzheimer's disease to ordinary aging, this exact kind of decision, keeping track of where you have and haven't already been, is one of the first things to falter. The radial arm maze has spent nearly fifty years turning that type of failure into something a researcher can actually measure.
A radial arm maze is a small circular hub with several arms, traditionally eight, branching outward like spokes, each ending in a spot where a food reward can be hidden. Olton and Samuelson designed it in 1976 to study how rats keep track of a shifting list of choices, what they called "serial learning." Almost fifty years on, the maze endures because it does something very useful: unlike a simpler open field or plus-maze, it forces an animal through a whole sequence of choices in a single trial, so one run yields data on many decisions instead of just one.
It also sits alongside the family of other spatial memory tests, including the Morris water maze, which asks an animal to find a hidden platform using distal cues, and the Y-maze, which asks it to alternate between three arms. Kohler and colleagues, comparing a classic radial maze to a newer, automated, food-deprivation-free version, note that the maze's core appeal is its ability to separate two memory systems within a single test, without requiring extensive pretraining [1].
The basic radial layout has even been adapted outside of memory research altogether. Ronquillo and colleagues built a three-dimensional radial maze, tracked with EthoVision, to study anxiety-like behavior instead, giving mice a continual choice between open, exposed paths and safer ones [2]. It's a reminder that the maze's real innovation isn't the food reward at the end of each arm, but the repeated, trackable choice in the middle.
Working memory and reference memory sound similar, but the radial arm maze can distinguish them both. Reference memory is an animal's stable knowledge of the task: which arms, out of all of them, are ever baited at all. Working memory is the moment-to-moment tracking within a single trial: which of those baited arms has already been visited. Because the maze can be configured either way, baiting every arm to test working memory alone, or baiting only some arms consistently to test reference memory as well, researchers can dial in exactly which kind of forgetting they want to study.
Does restoring the brain's own housekeeping chemistry help a failing memory? Gall and colleagues tested this in a streptozotocin rat model of sporadic Alzheimer's disease, a model that mimics some of the disease's brain-insulin-resistance features. Over 40 days, rats received melatonin, then ran a radial arm maze recorded and scored with EthoVision. Melatonin-treated rats made fewer working and reference memory errors than untreated model rats, and their brains showed less microglial and astrocyte activation, the hallmark scarring of neuroinflammation [3].
A separate line of work by Kim and colleagues has used the radial arm maze in 5xFAD mice, a strain engineered to overexpress mutant amyloid precursor protein. Across a series of studies, they tested microcurrent stimulation, alone and alongside the Alzheimer's drug donepezil, and reported improved radial maze performance alongside reduced amyloid plaque burden and neuroinflammation [4, 5]. In their most recent pilot combining microcurrent with donepezil, however, the combination only trended toward better performance than donepezil alone, a difference that did not reach statistical significance [6], a useful reminder that not every combination pans out, and that a null result is still a result worth reporting.
Does what an animal eats, or how inflamed its brain is, shape how well it navigates the maze? The same research group behind the melatonin study also asked whether restoring vitamin D could rescue memory in vitamin-D-deficient rats. Over several weeks, deficient rats ran a battery of tests, including the radial arm maze, tracked and analyzed with EthoVision. Vitamin-D-deficient rats performed worst in both the Morris water maze and the radial maze, while cholecalciferol-supplemented rats learned faster. The same rats, however, showed no difference on the novel object recognition test, a useful reminder that a treatment's benefits don't always generalize across every memory task [7].
A separate study, that also used EthoVision, looked at neuroinflammation more directly. Park and colleagues injected mice with lipopolysaccharide, a bacterial toxin that reliably triggers brain inflammation and behavioral impairment, then treated some of them with quinic acid, a plant-derived antioxidant compound. Mice that received quinic acid made fewer errors in the radial arm maze alongside lower markers of glial activation in the hippocampus [8].
Does removing a hunger signal in the brain change how well an animal remembers where the food actually is? Stark and colleagues silenced growth hormone secretagogue receptors, or GHSRs, the receptors that respond to the hunger hormone ghrelin, specifically in the olfactory bulb of mice. During the test, these mice made fewer arm entries and moved shorter distances overall, less motivated to forage. But when the researchers looked specifically at how often the mice chose the correct, rewarded arm over an incorrect one, there was no difference from control mice [9]. In other words, dulling an animal's drive to search did not dull its underlying spatial memory.
Does the same maze reveal memory loss differently depending on an animal's genes, or its sex? Borbely and colleagues compared young and aged mice lacking TRPA1, a receptor better known for sensing temperature and pain that also turns up in the brain. Aged wild-type mice made significantly more errors in the radial arm maze, and explored for longer before finding their rewards than young mice did. Aged TRPA1-knockout mice, by contrast, showed a smaller age-related decline. Notably, the same aged mice showed no significant memory difference on the Y-maze or Morris water maze, suggesting the radial arm maze picked up on an age effect that the other two tests missed entirely [10].
Szentes and colleagues found an equally nuanced picture studying mice lacking a somatostatin receptor subtype, sst4, recording every test, including the radial arm maze, with the same EthoVision-based setup. Deleting the receptor altered radial maze performance in female mice but not in males and aging itself improved working memory scores in females while leaving males unchanged [11].
Reproductive surgery is common, but does it change how the brain functions? Koebele and colleagues followed rats for up to a year after hysterectomy (the removal of the uterus), with or without ovary removal, testing them repeatedly in a water-based radial arm maze. Treated rats made more working memory errors than sham-operated rats, and this deficit was still present six weeks, seven months, and a full year after surgery [12].
Does learning something new physically change the brain's wiring, not just its chemistry? Shimizu and colleagues trained mice on an eight-arm radial maze whose doors opened and closed automatically, tracked and controlled from start to finish by EthoVision. Successful training stimulated the birth of new oligodendrocytes, the cells that wrap axons in insulating myelin, in the brain regions the task relies on most heavily, including the prefrontal cortex and hippocampus.
When the researchers genetically blocked the mice from making new myelin, radial maze performance suffered, and the mice that generated the most new oligodendrocytes during training were also the best performers [13]. It's a striking illustration of how far the maze's automation can go: in this study, the same EthoVision software that logged every arm entry was also programmed to operate the maze's doors.
Across all of these studies, the bottleneck is the same one a lot of these mazes have: dozens of trials generate a lot of individual decisions to score, and scoring them by hand is slow and prone to human error. As the oligodendrocyte study above shows, automation can extend well beyond scoring, into running the maze itself. With EthoVision, deep-learning detection tracks an animal's nose, body center, and tail base, so an arm entry, and a possible re-entry error, is logged the moment the animal genuinely commits to it. The software then reports arm sequences, latencies, and error counts automatically, trial after trial, without a researcher squinting at grainy footage.
The current release, EthoVision 19, adds a Noldus Calibration Card that sets up a maze's arena in seconds, along with refined AI tracking models that return data within moments of a recording finishing. For a maze built around one long list of possible errors, that kind of consistency is what turns a subtle memory deficit, or the absence of one, into something you can actually trust.
Fifty years after Olton and Samuelson built the first version, the radial arm maze is still finding new relevance: in Alzheimer's models, in diet and neuroinflammation, in feeding circuits, in genetics and sex differences, in myelin biology, and in the long cognitive tail of reproductive surgery. What ties these very different studies together is not the disease or the mechanism, but the maze's own design, forcing an animal to make one choice after another, so that memory becomes a sequence a researcher can count. Precise, automated counting is what lets researchers trust a subtle effect, or, just as importantly, an absent one.
For more on spatial memory testing, take a look at our recent blog on the Y-maze for learning and memory. If you would like to talk through how the radial arm maze, or automated tracking more broadly, could fit your own study, we are happy to help.
Bacterial infections in the brain can have lasting effects even after recovery. In this blog we will look a study that used PhenoTyper and the CognitionWall to study the mechanisms behind this cognitive impairment.
Read More arrow_forward
To improve standardisation and consistency in the novel object test, Spry et al designed the 'CapTouch' system: capacative objects that can automatically register interaction. But what about deep learning in EthoVision XT?
Read More arrow_forward
The Y-maze is one of the most efficient tools for studying spatial and working memory. Learn how automated tracking transforms arm entries into clean, objective data across rodent and zebrafish models.
Read More arrow_forwardWe'll get back to you shortly.
Please correct the following errors: