Method of Loci: Why Placing Beats Repeating

Introduction

Ask someone to recite twenty random words in order and they will fail. Ask them to describe the layout of the home they grew up in and they will keep going for ten minutes. The front door. The smell of the hallway. Which stair creaked. Where the coats hung.

That gap is the whole idea behind the method of loci. Memory for places is close to effortless. Memory for arbitrary lists is close to hopeless. So the technique takes the second thing and smuggles it inside the first. You pick a route you already know, you place a vivid image of each item at a fixed point along that route, and later you walk the route in your head and collect what you left there.

It works. A systematic review and meta-analysis published in the British Journal of Psychology in June 2025 found a large effect on immediate serial recall in adults compared with plain rehearsal, with a standardised difference of d = 0.88 and a 95% confidence interval running from 0.47 to 1.25 [1]. An earlier synthesis of 13 randomised controlled trials found a medium effect of g = 0.65, confidence interval 0.45 to 0.85 [2]. Two independent teams, two methods, same direction.

Now the part almost nobody writes down. In that 2025 review, roughly 80 percent of the studies feeding the effectiveness analysis were rated at high risk of bias, with a further slice rated serious [1]. The direction of the effect is solid. The exact size of it, in anything resembling real study conditions, is not.

This article is about both halves of that sentence. Where the technique came from, what the brain is doing when it works, what the strongest evidence actually supports, and the several popular claims about it that quietly fall apart when you check them.

The banquet that collapsed

Every account of this technique starts with the same story, and the story is almost certainly polished.

Cicero tells it in De Oratore, written in 55 BCE [3]. The poet Simonides of Ceos is performing at a banquet in Thessaly. He is called outside. While he is gone the roof comes down and kills everyone inside. The bodies are crushed beyond recognition and the families cannot identify their dead. Simonides can. He remembers where each guest had been sitting, and the seating plan gives him the names.

Whether any of that happened is unknowable. What matters is the conclusion Cicero draws from it, which is a genuine claim about cognition: order imposed by place makes recall possible.

The oldest surviving instruction manual is not Cicero. It is the anonymous Rhetorica ad Herennium, written around 90 BCE, which lays out the rules with startling practicality [4]. Choose backgrounds that are deserted rather than crowded. Space them so they do not blur into each other. Avoid places that are too bright or too dim. Make the images striking. Two thousand years before anyone measured a reaction time, someone had already worked out that distinctiveness and spacing matter.

Quintilian, writing around 95 CE, adds the detail most guides still repeat without knowing the source, which is the instruction to walk the same route in the same direction every time [5]. He was blunt about the limits too. He thought it worked well for lists and badly for continuous prose.

Then the technique goes quiet in popular memory and loud in scholarly practice. Frances Yates traced its survival through the medieval and Renaissance worlds in The Art of Memory, published in 1966, which is still the book that reopened the subject for modern readers [6]. Mary Carruthers went further into the monastic period and argued that memory craft was not a party trick in medieval intellectual life but the foundation of how reading, composition and prayer were actually done [7].

The strangest chapter is the Jesuit one. Matteo Ricci arrived in China in the late sixteenth century and taught the technique to Chinese scholars, publishing a treatise on it in Chinese in 1596. Jonathan Spence reconstructed that episode in The Memory Palace of Matteo Ricci [8]. Around the same period Giordano Bruno was pushing the art in the opposite direction, building elaborate symbolic memory systems in his 1582 work De umbris idearum that were less study tool and more cosmological machine.

Here is the sequence, compressed.

Notice the gap in the middle. For roughly nineteen centuries this was a craft passed between practitioners with no way to test whether the rules were true. The testing only starts once psychology has laboratories, and the neuroscience only starts once someone puts an electrode in a hippocampus.

What the numbers actually say

Most articles about this technique either quote nothing or quote something invented. One widely shared claim is that the method improves recall by two to three times. That figure has no traceable source in the research literature and should be treated as marketing rather than data.

Here is what the two syntheses report.

Both effects survive the usual stress tests. In the 2021 analysis the result held after adjusting for publication bias, after leave-one-out checks, after varying the control condition, and after removing outliers [2]. That is a reassuring pattern. An effect that only exists in one subset of studies usually collapses under those procedures.

The 2025 review is the more demanding of the two. It applied a formal risk of bias tool, used the GRADE approach for certainty, and ran a Bayesian model averaging analysis rather than a single fixed estimate. Its Bayes factor for the presence of an effect on immediate serial recall came out around 162, which is strong evidence that something real is happening [1]. The same paper is openly critical of the earlier meta-analysis for not separating younger from older adults, for not distinguishing studies with very different aims, and for treating different retention intervals as interchangeable.

Three limitations run through nearly all of this work and they matter more than the headline number.

Participants are overwhelmingly young psychology undergraduates. Training and testing usually happen on the same day. And the material being memorised is usually a word list, which is exactly the task the technique was designed for and almost nothing like a textbook chapter.

None of that makes the effect fake. It makes it narrow. The honest summary is that placing items along a known route reliably beats repeating them, for ordered material, in the short term, in conditions that have been measured well. Everything beyond that is extrapolation.

There is one older result worth putting next to these. In 1980, Roediger compared four mnemonic systems directly and found that the ones imposing an ordered spatial or sequential framework outperformed unstructured strategies for ordered recall [9]. The structure was doing the work, not the novelty.

The brain was already built for this

The reason a two thousand year old trick keeps working is that it was never really a trick. It borrows a system the brain maintains anyway.

In 1971, John O'Keefe and Jonathan Dostrovsky recorded from single neurons in the hippocampus of freely moving rats and found cells that fired when the animal was in one particular part of its environment and stayed quiet elsewhere [10]. They called them place cells. The hippocampus, it turned out, was not just storing memories in some general sense. It was holding a map.

Thirty four years later, Torkel Hafting and colleagues in the Moser lab found the layer underneath. In the entorhinal cortex, which feeds the hippocampus, certain neurons fire in a repeating triangular grid across an entire environment [11]. Grid cells are effectively a coordinate system. Place cells say where you are. Grid cells provide the metric.

In 2014 the Nobel Assembly awarded the prize in Physiology or Medicine to O'Keefe and to May-Britt and Edvard Moser for these discoveries, describing the result as an inner positioning system in the brain [12].

Now put the two facts side by side. The hippocampus runs a spatial map, and the hippocampus is also central to forming new episodic memories. That is not a coincidence of anatomy. The circuitry that binds "what" to "where" is the same circuitry that binds an item to a context, which is why the hippocampus decides what gets kept in the first place.

The method of loci exploits that overlap deliberately. It converts a list, which the brain has no evolved machinery for, into a set of locations, which the brain has dedicated hardware for. Related work on context-dependent memory shows the same principle running in the background of ordinary learning, where the place you studied quietly becomes part of the memory trace whether you intended it or not.

What happens inside a trained brain

If the technique really recruits spatial circuitry, trained users should look different under a scanner. They do, but not in the way people expect.

In 2003, Eleanor Maguire and colleagues scanned ten people with genuinely exceptional memory performance, most of them ranked competitors, and compared them with matched controls [13]. Nine of the ten used the method of loci. The finding that made the paper famous is what they did not find. No superior general intelligence. No structural brain differences. What differed was the pattern of activation during encoding, concentrated in regions tied to spatial memory and route learning, including the retrosplenial cortex and the right posterior hippocampus.

That result reframed the whole subject. Exceptional memory looked like a learned strategy running on ordinary hardware.

The 2017 study by Martin Dresler and colleagues tested that directly [14]. They scanned 23 of the world's top memory competitors, then took people with no training and put them through six weeks of method of loci practice, amounting to forty sessions of about half an hour. A second group did working memory training instead. A third did nothing.

The loci group more than doubled the number of words they could recall from a 72 item list. Neither control group came close. And the improvement was still measurable four months later.

The brain data was more interesting than the behaviour. Training did not grow a region. It rewired the connections between regions, shifting patterns of functional connectivity between visual, medial temporal and default mode networks toward the pattern seen in the competitors. Memory skill turned out to be a network property.

Isabella Wagner and colleagues followed this up in 2021 with a result that sounds backwards [15]. After mnemonic training, task related activity in lateral prefrontal, parahippocampal and retrosplenial cortices went down rather than up, in both trained novices and elite competitors. Lower activation was partly associated with better performance at the four month follow up. The trained brain was not working harder. It was working more cheaply, which is what expertise usually looks like once a skill consolidates.

A separate structural analysis from the same broader research group reported that in memory athletes the volumes of the right hippocampus and the caudate nucleus were more strongly correlated with each other than in controls, and that this coupling tracked competitive ranking [16]. The suggestion is that a memory system and a habit system learn to cooperate. The upper end is well documented. Rajveer Meena recalled 70,000 decimal places of pi blindfolded at VIT Vellore on 21 March 2015, taking nearly ten hours [43].

One more result deserves mention with a clear warning attached. A 2025 preprint led by Jing Ren, with Wagner and Dresler as senior authors, reports distinct neural representations in prefrontal, inferior temporal and posterior parietal regions in both memory athletes and a six week training group, and links how distinct those representations were to memory performance four months later [17]. This has not been peer reviewed and should be treated as provisional until it is.

Why placing beats repeating

The neuroscience explains where. The cognitive psychology explains why.

Fergus Craik and Robert Lockhart proposed in 1972 that memory strength depends less on how long you hold something and more on how deeply you process it [18]. Rehearsing a word by repeating its sound is shallow. Working out what it means, connecting it to something you already know, forming an image around it, all of that is deep. Deep processing wins, consistently.

The method of loci forces deep processing whether you want it or not. You cannot place an item at a locus without doing three things: retrieving the locus, generating an image, and binding the two together. That is elaborative encoding by construction, not by good intentions.

It adds a second thing that plain rehearsal cannot. Retrieval structure.

When you try to recall a list with no scaffold, you have to generate the items from nothing. When you have a route, you generate the locus instead, and the locus generates the item. You have swapped a hard free recall problem for an easy cued recall problem, repeated as many times as you have loci. That is the mechanism in one sentence.

The dependency in that diagram is the technique's weak point as well as its engine. If the route is shaky, nothing downstream survives. This is why the classical sources insisted on using places you already know cold, and why the modern advice to invent an imaginary palace from scratch is usually bad advice for beginners.

There is a useful piece of evidence about what expertise does here. In 1980, Anders Ericsson and colleagues reported a training study in which a single participant, starting from an ordinary digit span, reached a span of around eighty digits after extended practice by encoding digit groups into meaningful units [19]. He did not expand his working memory. He built retrieval structures in long term memory and pushed the load into them. Ericsson and Walter Kintsch later formalised that idea as long term working memory [20]. A memory palace is one instance of exactly that architecture. Anyone curious about how this generalises will find the same pattern in the psychology of expertise, where skilled performers consistently outperform not through faster hardware but through better organised knowledge.

Three layers people keep confusing

A lot of study advice collapses different problems into one bucket. It helps to separate three.

Encoding is getting the material in. The method of loci is an encoding technique. It changes the quality of the trace you lay down.

Retrieval practice is pulling material back out under effort. Testing yourself strengthens memory more than restudying does, a finding Henry Roediger and Jeffrey Karpicke demonstrated cleanly in 2006 and which has since been replicated across settings and materials [21]. If you want the full mechanism, retrieval practice and the brain covers it in depth.

Scheduling is deciding when that practice happens. Spreading study across time beats massing it into one block, an effect quantified across hundreds of comparisons by Nicholas Cepeda and colleagues in 2006 [22]. The background is in the spacing effect.

These three do different jobs and do not substitute for each other. A beautifully constructed palace that you never test yourself on will still fade. Perfectly spaced review of material you encoded carelessly will still be slow going. The classical sources understood the first half of this instinctively, which is why Quintilian insisted on walking the route repeatedly rather than building it once and walking away.

Worth saying plainly: the loci literature almost never runs the comparison people actually want, which is method of loci plus spaced retrieval against spaced retrieval alone. Until someone does that trial properly, claims about how the pieces stack remain reasoned inference rather than measured fact.

The parts the internet gets wrong

This is where the popular guides and the evidence part company.

The bizarreness one is worth dwelling on because it is repeated everywhere. Mark McDaniel and Gilles Einstein showed in 1986 that the effect is fragile and context dependent, showing up when bizarre and common items compete within the same list [23]. Later analysis found no bizarreness advantage when the two item types were recalled separately, pointing at retrieval order rather than encoding strength [24].

The practical translation is not "make images boring". It is that vividness and clarity matter, and grotesque novelty for its own sake is a cost you may be paying for nothing.

Does the route even matter?

This is the most interesting open question in the field and almost no popular guide mentions it exists.

Jeremy Caplan and colleagues at the University of Alberta built three virtual environments that differed sharply in how navigable they were: an ordinary apartment, an open field with no obvious path, and a radial maze. If the walking is doing the work, performance should track navigability. It barely did [25]. The authors suggested the method of loci may be better understood as a special case of imagery based peg systems, where what matters is having an ordered set of stable cues rather than a path connecting them.

The same group tested another version of the question by comparing different mnemonic scaffolds for serial recall, including one built on parts of the body rather than places in the world. The body based scaffold held up as a strong alternative to the classical route [26].

Take that seriously and the ancient instruction to walk a fixed path in a fixed direction may be an artefact of how the technique was taught, not a requirement of how it works. The pieces that survive scrutiny are the ones the Rhetorica ad Herennium already emphasised: a set of distinct, well spaced, already familiar anchors, used in a consistent order.

That is a smaller claim than "build a palace". It is also a more defensible one.

What about people who cannot picture anything

Aphantasia is the absence or near absence of voluntary visual imagery. Adam Zeman and colleagues described the condition formally in 2015 [28], and Carla Dance and colleagues later estimated its prevalence in the general population, finding figures that vary considerably depending on which cut off and which questionnaire is used [29].

The obvious worry is that a technique built on mental imagery should be useless for these people. The evidence does not support that worry as strongly as you would expect.

Two findings are worth separating. In work on interactive imagery for word pairs, aphantasic participants who reported little or no mental imagery benefited from imagery instructions as much as controls, and neither reported vividness nor objective imagery skill explained the benefit [27]. Separately, in a study comparing mnemonic scaffolds directly, measures of visual imagery aptitude and vividness did not predict accuracy for the method of loci itself [26]. Whatever the instruction is doing, it does not appear to depend on the subjective picture being bright.

One plausible reading is that the useful part of "imagine it" is relational rather than pictorial. Deciding that the item and the locus interact in some specific way may be the operation that matters, and the visual experience may be a side effect rather than the cause.

The honest position is that this question is under powered and unresolved. The aphantasia result above came from a paired associate task, not from a full memory palace, and nobody has run a large trial of the method of loci itself in people with severe aphantasia. Until that exists, "you need to be a visual person" should be treated as an untested assumption rather than a fact.

Where it works outside the laboratory

Word lists in a psychology building are one thing. Real material is another.

The most cited educational test is a study of medical students learning endocrinology, where a group that built memory palaces for insulin and diabetes content outperformed a group receiving lectures and self directed study, with the difference reported as highly significant [30]. It is a small, single institution, quasi experimental study and it should be read as encouraging rather than conclusive. Domains with heavy ordered visual content, such as anatomy, are the natural fit, and the reasoning behind that is covered in visual memory in anatomy.

Ageing is where the picture gets genuinely complicated.

Paul Verhaeghen and colleagues found in a 1992 meta-analysis that mnemonic training does improve memory performance in older adults [31]. Good news. But Lars Nyberg and colleagues added an uncomfortable finding in 2003: training helped both younger and older adults, and the gap between them widened rather than narrowed [32]. Older participants showed less of the frontal and occipito parietal recruitment that accompanied the gains in younger ones. Training raised everyone, and raised the young more.

There is a structural counterpoint. Andreas Engvig and colleagues reported cortical thickness changes in older adults following memory training in 2010 [33]. The ageing brain is not inert.

The clinical work is the part that gets scattered across sources and almost never assembled.

Tim Dalgleish and colleagues in 2013 trained people with depression, both in episode and in remission, to store self affirming personal memories using the method of loci, against a chunk and rehearse control [34]. Both groups performed near ceiling after a week. At a surprise follow up a week after that, recall held in the loci group and dropped in the rehearsal group. A follow up study by Aliza Werner-Seidler and Dalgleish extended the retention advantage to three months and reported that participants used their memory store spontaneously in daily life [35]. Participants worked with fifteen loci along a familiar journey.

In rehabilitation, imagery and context based memory protocols closely related to this family of techniques have been tested in randomised trials. Nancy Chiaravalloti and colleagues reported a trial in multiple sclerosis in 2013 [36] and a double blind, placebo controlled trial in moderate to severe traumatic brain injury with 69 participants, 35 of them in the treatment group, using a ten session protocol delivered over five weeks [37]. A multi site trial protocol targeting mild cognitive impairment was published in 2025, so the older adult clinical question is actively being answered rather than settled [38].

A note of restraint. These clinical protocols are structured, therapist delivered and spread across multiple sessions. That is a long way from reading a guide and imagining a shopping list in your hallway, and borrowing their effect sizes to describe casual self study would misread what was tested.

Virtual palaces

If the technique depends on richly known real places, a briefly viewed virtual environment should not work. It does.

Eric Legge and colleagues compared a virtual environment studied for only a few minutes against participants' own homes and found equivalent memory performance across ten lists of eleven words [39]. The title of the paper says it plainly: a memory palace can be built in minutes.

Immersion appears to add a modest amount. Eric Krokos, Catherine Plaisant and Amitabh Varshney asked forty participants to recall the placement of well known faces in two virtual environments, comparing a head tracked head mounted display against a desktop screen, and reported an overall improvement in recall accuracy of 8.8 percent for the immersive condition [40]. Statistically real, and considerably smaller than the enthusiasm around virtual memory training would suggest.

The interesting implication is not about hardware. It is that deep familiarity with a place may be less necessary than the tradition assumed. What seems to matter is that the anchors are distinct and stably ordered, not that you have lived among them for years.

Older than Greece

The Simonides story places the origin in fifth century BCE Greece. There is good reason to think spatial memory systems are far older and far more widespread than that.

Lynne Kelly has argued across several works, including a Cambridge University Press monograph on knowledge and power in prehistoric societies, that oral cultures worldwide developed place based memory systems capable of encoding very large bodies of practical knowledge [41]. The range is wide. Australian Aboriginal songlines link knowledge to sequences of terrain features across enormous distances, while the lukasa memory boards of the Luba people encode narrative and genealogical information in beads and carved surfaces. Pacific navigation traditions and ceremonial routes across the Americas show similar patterns.

Two things should be said carefully here.

The first is that this scholarship is well supported and was developed in collaboration with Aboriginal Elders and knowledge holders, which gives it a grounding that armchair reconstruction lacks. The second is that it remains partly interpretive. Reading a specific archaeological site as a memory device is an inference, not a measurement, and it has not reached the status of settled archaeology. Both statements can be true at once.

There is now peer reviewed experimental work in this space. Researchers led by David Reser translated Australian Aboriginal memory techniques into a medical education setting and compared them against method of loci style training and a control, publishing the results in PLOS ONE in 2021 [42]. Bringing an oral tradition into a randomised classroom study is unusual and valuable, because it tests the tradition rather than merely admiring it.

If the broad picture holds, the conclusion is not that Greece invented anything. It is that multiple cultures independently discovered the same fact about human cognition, which is roughly what you would expect if the fact is grounded in shared neural architecture rather than cultural transmission.

Where the method fails

Enthusiast guides rarely discuss failure. The failure modes are predictable and worth knowing before you invest hours in building something.

Interference from reuse is the most common. Palaces are tempting to reuse because building them is the expensive part. But putting a second set of items at the same loci creates exactly the conditions under which old associations intrude on new ones, a pattern documented across the interference literature and explained in proactive interference. Competitors handle this by maintaining large numbers of separate routes and by leaving time between reuses. Casual users usually do neither.

Decay of unrehearsed associations is the second. The durability results are genuinely encouraging, with gains still present at four months [14] and retention holding at three months in the clinical work [35]. But every one of those studies involved sustained practice. A palace built once and never revisited behaves like anything else built once and never revisited. The shape of that decline is the subject of the forgetting curve.

Abstract material resists. The classical sources already knew this and drew a line between memory for things and memory for words. Placing a concrete noun at a locus is straightforward. Placing a legal principle, a statistical concept or an argument requires you to invent a concrete stand in first, and the quality of that stand in determines everything. Quintilian's scepticism about using the method for continuous prose has aged well.

Encoding cost is real. Every item needs a locus retrieved, an image generated and a binding formed. For a twenty item list that is a fair trade. For several hundred facts it is a substantial time investment that competes directly with simply testing yourself on the material.

And there is a quieter failure that rarely gets named. The technique is very good at producing ordered recall of items, which can feel like mastery while leaving understanding untouched. Being able to list the stages of a process in order is not the same as knowing why the process works. For anything conceptual, the palace is scaffolding, not the building.

What the evidence supports, and what it does not

The honest verdict

The method of loci is one of the few study techniques with a documented history stretching back two millennia and a modern effect size attached to it. That combination is rare and it is why the technique deserves the attention it gets.

It is also surrounded by more confident claims than the data can carry.

What holds up: placing items at ordered, distinct, familiar anchors reliably beats repeating them, for material that has an order, and the advantage persists for months when practice continues. The mechanism is not mysterious. The technique borrows the hippocampal and entorhinal machinery the brain maintains for knowing where things are, and converts a task the brain is bad at into one it is good at. Trained brains reorganise their connectivity rather than growing new regions, and after enough practice they do the same work with less effort.

What does not hold up: the specific route, the requirement for bizarre imagery, the assumption that you need vivid visual imagination, and the implication that this is a general upgrade to learning. It is not a general upgrade. It is a very good solution to a specific problem, which is ordered recall of discrete items.

And the caveat that should follow the headline number everywhere it appears. The 2025 estimate of d = 0.88 sits on a literature where about 80 percent of the contributing studies were rated at high risk of bias, where participants are mostly young students, and where testing usually happens on the same day as training [1]. The direction is trustworthy. The magnitude, applied to your own textbook chapter next month, is a guess.

Simonides, if the story is true, worked out something real at a funeral. Two thousand years later the measurement is finally catching up with the intuition, and the measurement is more modest and more interesting than the legend.