Friday, 25 November 2016

Your dog remembers more than you think

At last, scientists may have an answer to a question every dog owner asks: Does your pet remember the things you do together? For people, at least, the ability to consciously recall personal experiences and events is thought to be linked to self-awareness. It shapes how we think about the past—and how we predict the future. Now, a new study suggests that dogs also have this type of memory, indicating that the talent may be more common in other animals than previously recognized.
The study, “is a creative approach to trying to capture what’s on a dog’s mind,” says Alexandra Horowitz, a dog cognition scientist at Barnard College in New York City who was not involved in the research.
The idea that nonhuman animals can consciously remember things they’ve done or witnessed in the past, called episodic memory, is controversial—largely because it’s thought that these animals aren’t self-aware. But scientists have shown that species like Western scrub jays, hummingbirds, rats, and the great apes—those that have to recall complex sequences of information in order to survive—have “episodiclike” memory. For instance, the jays remember what food they’ve hidden, where they stashed it, when they did so, and who was watching while they did it.
But what about recalling things that aren’t strictly necessary for survival, or someone else’s actions? To find out whether dogs can remember such details, scientists asked 17 owners to teach their pets a trick called “do as I do.” The dogs learned, for instance, that after watching their owner jump in the air, they should do the same when commanded to “do it!”
<p>After watching his owner touch an umbrella, a dog imitates the action. &nbsp;</p>
After watching his owner touch an umbrella, a dog imitates the action.  
Mirko Lui
“But that alone doesn’t prove episodiclike memory,” says Claudia Fugazza, an ethologist at Eötvös Loránd University in Budapest, and the study’s lead author. “You have to test them when they don’t expect it.”
So in the next round of training, the owners taught their dogs to lie down after watching them do something like touching an open umbrella or stepping up on a chair. They were no longer required to imitate. “And then we surprised them,” Fugazza says. Again, an owner performed an action, but this time after the dog laid down, the owner ordered, “Do it!” The dog then had to recall what it had seen its owner do, even though it had no expectation that it needed to remember the action. The dogs were tested in this way both 1 minute and 1 hour after watching their owners.
The dogs succeeded in 33 of 35 trials. That suggests that dogs have something similar to episodic memory, Fugazza and her team report today in Current Biology. But the longer the canines wait, the more trouble they have recalling the action. That’s similar to human episodic memory, which decays at a faster rate when an event isn’t intentionally recorded, the researchers say. For instance, you’re more likely to remember your first kiss than a hug last week from your spouse.
“It shows that our dogs remember events much like we do, and [it] blows out of the water the old way that most scientists would characterize animal memory,” says Brian Hare, a dog cognition expert at Duke University in Durham, North Carolina, who was not involved in the study. “Our dogs’ memories aren’t based simply on repetition and reward.”
Discovering that this type of memory is not unique to humans means it “did not evolve only in primates, but is a more widespread skill in the animal kingdom,” Fugazza says. Indeed, parrots, dolphins, and killer whales could be the next ones to be tested, because other researchers have already taught these species to “do as I do.” Most likely, they remember more than we think.

Thumbnail image of Figure 1. Opens large image

Figure 1

Duration of Looking at Owner
Duration of looking at the owner in Do as I Do tests of dogs after expected (white bars) or unexpected (gray bars) “Do it!” commands were given. Bars with continuous lines represent data from the present study; bars with dashed lines represent data from two previous studies with similar conditions but expected recall (1 min retention time with “Lie down” distraction before imitation[14]; 1 hr retention time[15]).
See also “Violation of expectation: dogs look longer at the owner if the test is unexpected” in Supplemental Experimental Procedures.

Expand allCollapse all


  • Episodic-like memory of others' actions was tested in dogs
  • Dogs were trained to imitate human actions with the Do as I Do method
  • Dogs could recall the owners' actions when unexpectedly requested to imitate
  • Memory of owners' actions decreased faster with increased test delay


The existence of episodic memory in non-human animals is a debated topic that has been investigated using different methodologies that reflect diverse theoretical approaches to its definition. A fundamental feature of episodic memory is recalling after incidental encoding, which can be assessed if the recall test is unexpected [1]. We used a modified version of the “Do as I Do” method [2], relying on dogs’ ability to imitate human actions, to test whether dogs can rely on episodic memory when recalling others’ actions from the past. Dogs were first trained to imitate human actions on command. Next, they were trained to perform a simple training exercise (lying down), irrespective of the previously demonstrated action. This way, we substituted their expectation to be required to imitate with the expectation to be required to lie down. We then tested whether dogs recalled the demonstrated actions by unexpectedly giving them the command to imitate, instead of lying down. Dogs were tested with a short (1 min) and a long (1 hr) retention interval. They were able to recall the demonstrated actions after both intervals; however, their performance declined more with time compared to conditions in which imitation was expected. These findings show that dogs recall past events as complex as human actions even if they do not expect the memory test, providing evidence for episodic-like memory. Dogs offer an ideal model to study episodic memory in non-human species, and this methodological approach allows investigating memory of complex, context-rich events.

Results and Discussion

Episodic memory has been defined as memory of personal events and specific episodes in one’s life, and it is thought to be linked to self-awareness [e.g., 3, 4]. Whether non-human animals possess some forms of episodic memory is a controversial topic, and it is difficult to design experimental procedures to assess self-awareness unambiguously. Therefore, this form of memory in non-human animals is referred to as “episodic-like memory.” The diversity of methods to investigate episodic-like memory reflects the controversies regarding its definition [5, 6, 7, 8]; however, recent approaches seem to agree that the recalling of an event relies on episodic memory when encoding of such event was incidental [9, 10, 11, 12, 13]. Incidental encoding occurs when information is stored without knowing that it has to be remembered or that it will be important later [9]. This requisite ensures that the subject cannot rely on learned rules (semantic memory) to succeed in the subsequent memory test. Because, at present, no experimental procedure exists to directly assess the type of encoding (i.e., a subjective state), a crucial criterion of studies focusing on episodic-like memory is that the recall test should be unexpected [1]. Unexpectedness of the test ensures that there is no specific motivation for explicit encoding, so incidental encoding can be reasonably assumed.
To explore the ability of dogs to recall past events when there was no expectation of the recall test, we used an innovative methodology: a modified version of the “Do as I Do” paradigm, relying on dogs’ ability to imitate human actions after a delay [14, 15]. Our aim was to test dogs’ episodic-like memory of past events (i.e., human actions) that are richer in content and more complex than what was tested in the majority of previous studies [e.g., 7, 9, 16, 17]. In most of these studies, laboratory animals were tested on memory of simple events, such as object exploration or feedings. Although these findings provide important advances for the study of episodic memory, real-life events are far more complex and richer in content. Particularly, from a pet dog’s perspective, the actions of humans are arbitrary behaviors that are always potentially different and can be performed on many different objects and in many different contexts. Episodic-like memory of such context-rich events was not tested previously in non-human species, except for chimpanzees and orangutans [5]; thus, it is not known whether this ability evolved only in primates or is a more widespread trait.
Here, we investigate whether dogs can rely on episodic-like memory to recall context-rich events from the past. We hypothesized that dogs can rely on episodic-like-memory to recall and imitate incidentally encoded actions performed by their owners, and we tested two predictions. First, we expected dogs to be able to imitate incidentally encoded actions when the imitation test was unexpected, albeit less successfully compared to their baseline imitation success when recall is expected. Second, we predicted that imitation success would decrease significantly with longer retention intervals, as memory appears to decay faster when encoding is incidental as opposed to when it is intentional [7, 18, 19]. Before testing, pet dogs were trained in two stages; the first stage has been independent of this study, as we enrolled dogs that were previously trained by their owners with the regular Do as I Do training to imitate human actions on command “Do it!” [2, 14] (for more details, see “Do as I Do training” in Supplemental Experimental Procedures). At the beginning of this study, the dogs’ baseline imitation success was assessed with the two-action method [20] in an expected imitation test (“baseline imitation” henceforth). Every dog had been exposed once to the demonstration of one of two possible novel (not trained) actions on an object (e.g., climb on a chair or touch the chair with paws; see Table S1). After the demonstration, the owner gave the “Do it!” command. Dogs were then free to perform any action, including other actions than those chosen for the tests.
To ensure that the subsequent imitation test was unexpected, after the baseline test, dogs underwent a second stage of training in which they were not required to imitate anymore. Instead, after the owners’ demonstration of various actions in sessions of six different trials, dogs were always required to perform a simple training exercise: lying down (“Lie down training” in Supplemental Experimental Procedures). The aim of this training was to substitute the dogs’ expectation of the imitation command with the expectation of a “Lie down” command. After the successful “Lie down” training, we tested dogs’ memory of unfamiliar (previously not trained or tested) actions by unexpectedly commanding them to imitate instead of lying down (Movie S1). Dogs were tested with the “Do it!” command only if they lied down spontaneously after the demonstration, suggesting with their behavior that they expected a lie down command (all dogs lied down spontaneously).
Dogs were not allowed to motor practice the demonstrated actions; they could only observe them during the demonstration. We tested each dog in two imitation tests: after retention intervals (i.e., time between demonstration and the “Do it!” command to imitate) of 1 min and 1 hr, in random order of the delays and the demonstrated actions. The tests were video recorded and later behaviorally coded for statistical analysis. It is reasonable to assume that in these tests, a successful imitation of the previously demonstrated action was possible only if dogs encoded the action incidentally, because the imitation test was unexpected, so there was no motivation for the dogs to encode them explicitly. Therefore, our method complies with the requirements for testing episodic memory (recall of an incidentally encoded event assessed by an unexpected recall test).
The unexpectedness of the test is a critical and at the same time challenging issue because it is difficult to assess the mental state of non-verbal subjects (i.e., acquire information about their expectations). Previous studies relied on the mere assumption that the test was unexpected [5, 9, 10, 11, 12, 13]. In contrast, we experimentally modified dogs’ expectations and searched for behavioral evidence for this. First, we ensured that the dogs expected to receive the “Lie down” command—and not the imitation command—by training all dogs until they spontaneously lied down after they had seen the demonstrated actions in at least five of six trials in two consecutive training sessions. In the unexpected tests, all dogs lied down spontaneously after the demonstrated actions, indicating that they expected a “Lie down” command, not an imitation command. Second, we relied on the well-established violation of expectation paradigm [e.g., 21, 22] that has also been successfully used in dogs [23, 24, 25]. This paradigm predicts a longer duration of looking toward the source of violation of expectation; therefore, we expected longer duration of looking at the owner who issued the “Do it!” command when this was unexpected as opposed to when it was expected. Because of the excess of zeros in the expected imitation tests (due to dogs that did not look at the owner after the “Do it!” command was given), we analyzed duration of looking in Tweedie Generalized Linear Mixed Models (GLMMs; package “cplm” [26] in R statistical environment, v. 3.2.3 [27]), with dog ID as random term and test condition as fixed effect (factor with three levels: baseline [expected imitation test], 1 min [unexpected imitation test], and 1 hr [unexpected imitation test]). Dogs looked significantly longer at the owner in the two conditions with unexpected imitation test than in the baseline condition with expected imitation test (likelihood ratio test of Tweedie GLMMs with and without test condition as fixed factor: χ22 = 25.45, p < 0.001; Figure 1). Other than expectedness, longer duration of looking may be explained by between-group differences in retention times and the effect of distraction by the “Lie down” command. Therefore, we excluded these alternative explanations in further analyses comparing the duration of looking in the present study with that of previous studies with identical delays but expected imitation tests (Figure 1 and “Violation of expectation” analysis in Supplemental Experimental Procedures).
Imitation success (binary response variable) was analyzed using binomial GLMMs (R package “lme4” [27]), with dog ID as random term and test condition as fixed effect (factor with three levels: baseline [expected imitation test], 1 min [unexpected imitation test], and 1 hr [unexpected imitation test]). In support of both of our predictions, we found that dogs were able to imitate when the imitation test was unexpected (although less successfully than when it was expected), and imitation success decreased quickly (i.e., fewer subjects imitated) with increasing retention interval (GLMM of imitation success, effect of test condition: χ22 = 14.7, p < 0.001; Table 1; Figure 2). A more rapid decay of dogs’ memory as a result of incidental encoding was apparent when we compared imitation success after 1 min and 1 hr retention interval when recalling was unexpected (this study) with results of our previous studies, with similar conditions (also with 1 min and 1 hr retention intervals) but when the imitation test was expected [14, 15] (Figure 2). When the recall test was expected, imitation success of dogs was not significantly different between immediate recall and recall after 1 hr delay [15]. In addition, imitation success with expected recall was more than 2-fold compared to when recall was unexpected (binomial GLM of imitation success after 1 hr retention intervals, expected recall [from 15] versus unexpected recall [this study]: 83.3% versus 35.3%; χ21 = 7.0, p = 0.008, regression coefficient [B ± SE] for expectedness = 2.22 ± 0.93).
Table 1Imitation Success
Effects of Test ConditionParameter Estimate ± SEzp
Intercept (baseline, expected)2.87 ± 1.112.590.010
Baseline (expected) → 1 min (unexpected)−2.50 ± 1.19−2.100.036
Baseline (expected) → 1 hr (unexpected)−3.51 ± 1.27−2.770.006
Imitation success (binary response variable) in Do as I Do tests of dogs (N = 17) based on whether recalling is expected and the length of retention interval. Parameter estimates with standard error (SE) between levels of test condition (fixed factor) and statistical significance are given from the binomial GLMM. Dogs were repeatedly tested in three test conditions: in baseline (expected imitation test) and after 1 min and 1 hr retention intervals (unexpected imitation tests), separately.
We argue that the difference in memory decay between this study and the previous one with identical delay [15] further corroborates that the dogs relied on an episodic-like memory in the present study, as this type of memory is proposed to decay faster with time than other types of long-term memory [18, 19]. In the case of expected imitation tests, dogs may have encoded the owners’ demonstrated actions explicitly because, as a result of previous training, they expected to be required to imitate. This implies that dogs might have used semantic memory to succeed in the deferred imitation task. In contrast, in the present study dogs were tested in the deferred imitation test only after assuring that their expectation of the future action required from them was different from the demonstrated action. Despite this, our results suggest that dogs could encode the demonstrated actions incidentally, although less successfully compared to the baseline.
Ostensive signals used by the owners to prevent dogs from moving during the demonstrations (“Stay and pay attention” command) may have increased the dogs’ attention, but this is unlikely to have resulted in using explicit memory in the unexpected tests. The same cues were also used during the “Lie down” training, in which dogs specifically learned that the owner’s subsequent actions were irrelevant. In addition, this command is commonly used in everyday life situations with pet dogs, whenever owners want to prevent their dogs from interfering with their activities. Following the concept of incidental encoding (not knowing that the information will be important later [1, 9]), we experimentally modified dogs’ expectations so that recalling the previously demonstrated actions was unexpectedly required. Although we provided multiple, independent experimental evidence for unexpectedness of recall (spontaneous lying down at the beginning of the test and behavioral signs of violation of expectation when unexpectedly required to imitate), we acknowledge that ensuring incidental encoding by direct evidence is problematic because it concerns the inner state of the subjects. Such direct and exclusive evidence seems extremely challenging to provide (if not impossible), so we relied on the assumption that the “Lie down” training resulted in dogs not explicitly encoding the demonstrated actions because these were irrelevant for the subsequent task. A steeper decrease in imitation success, albeit as an indirect evidence, strongly supports that we succeeded in this [18, 19].
Importantly, by using the two-action procedure in which two actions (A or B) are demonstrated on an object, our study provides evidence that the underlying process resulting in dogs’ reproduction of the demonstrated actions was deferred imitation (in 94.3% of all the tests when dogs performed action A or B, it was in correspondence with the demonstrated actions; Table 2; see Supplemental Experimental Procedures for more details on this analysis). This supports the notion that the dogs could imitate owners’ actions that were incidentally encoded without being presented with samples of those at the time of recall and without motor practicing during the retention interval.
Table 2Imitation Analysis
Test Condition and Retention IntervalDemonstrated ADemonstrated B
Performed APerformed BPerformed OtherPerformed APerformed BPerformed Other
Baseline imitation901070
1 min unexpected512063
1 hr unexpected505115
Number of dogs that performed action A, action B, or any other actions based on the demonstrated action in the various conditions of the Do as I Do test.
Testing for deferred imitation is a widely used approach to investigate the development of cognitive abilities in human infants [e.g., 28] and chimpanzees [29, 30]. These studies, however, were not specifically designed to investigate episodic memory, and it cannot be determined whether encoding of the demonstrations was incidental. Incidental encoding may also occur in cases of latent learning [31], although it has to be confirmed. Important advances about recall of incidentally acquired information were recently made by authors applying methods that rely on the unexpectedness of the recall test [9, 10, 11, 12, 13]. Zentall et al. [11] argued that in order to investigate episodic-like memory in non-verbal species, it is possible to teach them to use a trained behavioral response to “answer” a question about a past event (e.g., “Did you peck or not?”). Then the subjects can be “asked” this question unexpectedly, to assess whether they can remember the event. Using this method, the authors provided evidence that pigeons recall a simple species-specific action (pecking) and its location [12] after short delays. Zhou et al. [9] revealed that rats could not solve an unexpected memory task when the CA3 region of their hippocampus was inactivated, suggesting that this brain region is involved when recalling from memory is unexpected. Martin-Ordas et al. [5] tested chimpanzees and orangutans on their ability to recall the location of tools that they used previously to retrieve food. This study showed the ability to recall tool locations for long delays—even 3 years—after having used them. Although this suggests that some non-human species may recall events with a more complex nature than those tested in previous studies, the role of previous motor practice cannot be completely excluded due to the fact that those subjects performed the actions before testing. Mercado et al. [32] tested dolphins on their ability to reproduce the action they had just performed. Although, given the short delay, the subjects could have relied on their working memory, this methodological approach has the potential to test episodic-like memory for complex past events (one’s own actions) if subjects are prevented from keeping their mind actively on the actions so that the unexpectedness of the test can be ensured.
Our study makes an important advance in the study of episodic-like memory for multiple reasons. To our knowledge, this is the first time that a non-human species shows evidence of being able to recall complex events (i.e., others’ actions) without motor practicing on them during the retention interval—thus relying on a mental representation of the action that has been formed during incidental encoding, as assessed by an unexpected test. Note that in most previous studies of episodic-like memory, subjects participated in sample trials in which the same stimuli were presented as in test trials [e.g., 16, 17]. Our experimental procedure ensured that even if dogs were presented at the time of the test with the same objects that were used at the time of encoding, the specific actions performed by the demonstrator could only be imitated if dogs recalled a mental representation that was formed during encoding.
This modified version of the Do as I Do method has the potential to be applicable to a variety of species; the list of species in which the Do as I Do method has been used successfully includes dolphins (Tursiops truncatus) [33], parrots (Psittaciformes) [34], and killer whales (Orcinus orca) [35].
Moreover, to our knowledge this is the first study that experimentally addressed and behaviorally confirmed unexpectedness of the recall test. We believe that our research approach of modified expectation combined with the violation of expectation paradigm can be adapted to various experimental designs.
In conclusion, by using a modified version of the Do as I Do method, we found evidence that dogs can remember events as complex as human actions after incidental encoding, as assessed by an unexpected memory test, without motor practicing the actions during the retention interval and without being presented at the time of the memory test with the same samples presented when encoding took place. This is the first evidence of episodic-like memory of others’ actions in a non-human species, and it is the first report of this type of memory in dogs. We suggest that dogs might provide a new non-human animal model to study the complexity of incidental encoding of context-rich events, especially because of their evolutionary and developmental advantage to live in human social groups.

Author Contributions

The study was conceived by C.F., A.P., and A.M. The experiments were run by C.F.; A.P. analyzed the data. The article was drafted by C.F. and revised by A.P and A.M. All authors gave final approval for publication and agree to be held accountable for this work.


C.F. and Á.M. received funding from MTA-ELTE Comparative Ethology Research Group (MTA01 031). This project received funding from the Hungarian Scientific Research Fund (OTKA K109337). We are grateful to the dog owners who enthusiastically participated in this research with their dogs.

Supplemental Information

Movie S1. Related to Experimental Procedures
The video shows an unexpected imitation test after 1 min retention interval. A dog that was first trained with the Do as I Do method and then to lie down instead of imitate is unexpectedly commanded to imitate in the test.