Retrieval and Reconstruction

Assuming adequate consolidation (whatever that is), an encoded memory, made rich and distinctive by elaborative and organizational processing, remains permanently in memory storage until needed, at which time it must be retrieved from storage, subject to interference, and put to use.

Availability and Accessibility

The encoding and storage of a memory trace does not guarantee its later retrieval. Memories can be retained in storage even if they can't be retrieved on some particular attempt.

As discussed in the lectures on Storage, traumatic retrograde amnesia initially covers a relatively large swath of time before the injury occurred; but this very extensive RA typically remits to leave a more circumscribed "final RA" covering the moments to immediately prior to the accident.
The same pattern occurs in ECT-induced amnesia.
And in ECS-induced amnesia in rats, memories apparently obliterated by the ECS can be restored, to some extent, by reminder experiences.

These observations underscore a distinction, first articulated by Tulving and Pearlstone (1966), between the availability and the accessibility of a memory. In these instances, the RA has prevented access to memories that evidently remain available in storage -- because these memories can be recovered later on.

But you don't have to be amnesic to show the difference between availability and accessibility.

Consider, for example, what happens if an experimenter gives multiple memory tests after a single study trial. Typically, the number of items recalled will remain constant (unless there has been some distraction promoting forgetting over the retention interval). But, as Tulving (1964) observed, the fate of individual items varies considerably.

Some items are subject to inter-trial forgetting, meaning that they were remembered on an early trials, but forgotten on later ones.
Other items are subject to inter-trial recovery, meaning that they were forgotten on early trials, but remembered on later ones.

More than 50 years before Tulving, Ballard (1913) drew a similar contrast between two opposing memory processes:

Oblivescence, of the forgetting of what once had been remembered;
and reminiscence, or the remembering of what once had been forgotten.

In the usual case, as in Ebbinghaus's famous forgetting curve, oblivescence, or inter-trial forgetting, exceeds reminiscence, or inter-trial recovery. That does not always occur, but the fact that any reminiscence occurs at all illustrates the distinction between the availability of memory in storage and its accessibility on any particular retrieval attempt.

Ballard's distinction between oblivescence and reminiscence is illustrated by an experiment by Waldfogel (1948), who asked his college students simply to write down everything they could remember from the first 8 years of their lives - -giving them an hour to do so. The distribution of early childhood memories was markedly skewed, with most memories from age 6 and later -- a clear example of what is called infantile and childhood amnesia (discussed in the lectures on Memory Development).

But that's not all. One week later, Waldfogel repeated the task with the same subjects. This time, he got a slight increase in recall, meaning that some events were remembered on Test 2 that had not been remembered on Test 1. In fact, only about half of the events remembered on Test 2 had been recorded on Test 1 as well -- a dramatic illustration that the accessibility of available memories can fluctuate from trial to trial.

In some cases, as in Tulving (1964), inter-trial forgetting and inter-trial recovery cancel each other out, but circumstances, inter-trial recovery can exceed inter-trial forgetting, resulting in a net increase of recall -- exactly the opposite of Ebbinghaus's forgetting curve. Matthew Erdelyi (1978) has named this phenomenon hypermnesia, to contrast with amnesia. Erdelyi initially claimed that hypermnesia occurred with pictorial as opposed to verbal stimuli, but it now seems that the important variables have to do with elaboration and (perhaps) organization: anything that enhances encoding will reduce inter-trial forgetting, and foster inter-trial recovery -- facilitating access to information that has been available in memory all along.

Free Recall, Cued Recall, and Recognition

The distinction between availability and accessibility can also be observed in the feeling of knowing -- the familiar experience where you cannot recall a name, place, or word, but predict that you will recognize it when you see or hear it. These predictions are often correct. The fact that a memory was inaccessible on a recall test, but accessible on a recognition test, shows again that the memory was available all the time. (The related tip-of-the-tongue (TOT) state, in which people who can't recall a word can accurately describe some of its features even though they cannot recall the word itself, is another interesting phenomenon of metamemory.)

The superiority of recognition to recall has been known at least since McDougall (1924), leading to the general view that recognition is an "easier" test of memory than recall, just as multiple-choice tests are usually easier than short essays. But it is unclear just why this is the case -- what makes recognition easier than recall? This was not a problem that researchers studied during the heyday of interference theory, whose experiments were dominated by paired-associate learning (or, if you will, cued recall). But came to be of interest as a result of the cognitive revolution in memory.

Free Recall and Cued Recall

In an early experiment by Tulving and Pearlstone (1966), subjects heard lists of 12, 24, or 48 words representing 1, 2, or 4 conceptual categories. One group might hear 12 words, all from the same category, while another group might hear 48 words, 12 from each of 4 categories. You get the idea. In any case, the relevant category name was announced before the items themselves, and the items were also blocked by category. The subjects were then divided into two different tests.

Some subjects received a test of free recall, given 2 minutes to recall all the items they could.
Other subjects were given a test of category cued recall, where they were prompted with the category labels, and then asked to recall the relevant instances from the study list.

The basic result of the experiment was that cued recall was always superior to free recall, especially for longer lists. No particular surprise there, but after the conclusion of the free recall test, those subjects were given a second, cued recall test, on which they typically remembered additional items. Interestingly, most of the newly recalled items represented new categories, meaning that there were whole categories of items unrecalled on the initial, free recall test (and implying that if subjects recalled the category labels on the free recall test, they also pretty much exhausted their memory for those category instances).

It is important to understand that the subjects in this experiment had been treated identically up to the moment of the first memory test -- there were no differences in the conditions of encoding and storage conditions between the free-recall and category-cued-recall groups. Thus, the two groups were equivalent in terms of what was available in memory by virtue of encoding and storage processes. The differences in performance on the memory test were due to differences in retrieval conditions -- namely, that a cued recall test increases the accessibility of information stored in memory.

Note, too, that, cued recall is not just "easier" than free recall, because the relationship between test performance and the number of items per category differs between the two testing conditions.

Free recall performance was better when there were more items per category.
Cued recall performance was better when there were fewer items per category.

It was on the basis of these results that Tulving and Pearlstone (1966) drew their formal distinction between availability and accessibility. In their view, encoding and storage processes make information available in memory, while retrieval processes make information accessible in memory.

Any particular test of memory measures only that information which is accessible under the conditions prevailing at the time of retrieval.

Accessibility is clearly affected by encoding factors. elaboration and organization make information highly accessible; when encoding is impoverished, information is typically inaccessible

But accessibility is also affected by retrieval factors, particularly the cues provided at the time of retrieval. In free recall, the cues are relatively impoverished, while in cued recall, the cues are somewhat richer.

Similar observations were made in an experiment on retroactive inhibition by Tulving and Psotka (1971). In their experiment, subjects studied lists of 24 words, 4 items from each of 6 conceptual categories, and were given 3 study trials per list. Different groups of subjects memorized 1 to 6 lists. At the end of each study trial, the subjects were given a free recall test, providing a measure of learning during the study phase. And at the end of the last list, subjects were asked to recall all items from all lists. This free recall test was followed by a cued recall tests for the same items, in which subjects were presented with the category labels, and asked to recall the associated list items. Initial learning was pretty good, with subjects typically recalling about 75% of list items. But the final free recall test showed retroactive interference: Recall of items from a particular list decreased with the number of interpolated lists presented between the final study trial and the final test: this, of course is RI. But there was no evidence of RI with the cued recall test: the levels of cued recall were constant, regardless of the number of interpolated lists.

Thus, again, free recall (FRl) and cued recall (CRl) appear to differ qualitatively, not just quantitatively:

FRl and CRl show different relations between accessibility and category size;
FRl and CRl show different effects of retroactive interference.

In the previous studies, the list items were accompanied by conceptual category labels at the time of study, but this isn't necessary to see the effects of retrieval cues on memory.

In an experiment by Watkins and Tulving (1975), subjects studied lists of paired associates where the association between elements was either semantic (e.g., bark-dog) or phonemic (e.g., worse-nurse). As is usually the case in paired-associate learning, the first element in each pair is an explicit cue for the retrieval of the second element as a target.

On an initial cued recall test, the subjects were presented with the nominal cue from the study list, and asked to recall the targets. Cued recall was quite good, even though none of the cue-target relationships were categorical in nature.
On a second cued recall test, the subjects were presented with new retrieval cues that had not been presented during the study phase. These new cues reversed the initial relation between cue and target, such that an earlier semantic relation now became phonemic (e.g., grog-dog), and an initial phonemic relation now became semantic (e.g., doctor-nurse). These new cues retrieved items that had been forgotten on the earlier test.

True, semantic cues were generally better than phonemic cues, in that they retrieved more items on the initial test, and added more items on the second test. But it is also important to note that the cues presented on the second test were entirely new. Even though they had not been encoded at the time of study, they were still effective in gaining access to previously forgotten memories.

Recognition vs. Recall

Similar findings were obtained when recognition was added to the comparison. Tulving and Watkins (1975) asked subjects to study lists of 28 five-letter words, followed by one of the following tests:

Free recall, in which no (0) cues were provided to the subjects.
Cued recall, in which different groups of subjects were provided with the first 1-4 letters of each word as a retrieval cue.
Recognition, in which subjects were provided with all 5 letters.
Of course, subjects were provided with 0-5 letter cues for lure items as well.

The result was that cued recall was, as usual, better than free recall: Recall improves with the increasing number of cues. And recognition was best of all, presumably because a recognition test provides a copy cue -- a cue that is actually a copy of a studied item.

Light (1972) also performed a comparison of the effects of different types of retrieval cues. Her subject studied lists of words presented either alone or as part of meaningful sentences. Otherwise, no nominal cues were provided at the time of encoding.

In free recall, subjects were given no cue information.
In cued recall, subjects were told that the cues were synonyms of targets -- that is, words with similar meanings (e.g., house as a cue for home). Or, they were told that the cues were homonyms of the targets -- that is, words with the same spelling, but different meaning (e.g., quail meaning the bird, as a cue for quail, meaning to shake).
In recognition, the subjects were told that the cues were actual copies of studied items.

The result was that free recall yielded the worst performance, and recognition the best.

Homonyms were better cues than synonyms, which shows that cues based on semantic relations are not always the best.
But, interestingly, the homonym cues were worse than recognition -- even though both homonym cues and copy cues are both, in a sense, copies of the target (the same in spelling and punctuation)!

A Dual-Process Theory of Retrieval -- and a Single-Process Alternative

Thus, no matter how you slice it, Free Recall < Cued Recall < Recognition. One way to explain this outcome is in terms of a dual-process theory of retrieval, such as that offered by Anderson & Bower. In their theory,

Retrieval begins with a search process that generates candidate items in memory.
This is followed by a decision process that selects the appropriate items from among the candidates.

In the dual-process theory:

Recall requires both search and decision.
Recognition requires only decision, because the search process is obviated by the copy cues.

The dual-process theory of retrieval is supported by a number of observations:

First, of course, recognition (which requires only one process) is superior to recall (which requires two processes).
Stimulus frequency has different effects on recall and recognition: Recall favors high-frequency events (which are easy to locate during search), while recognition favors low-frequency events (which stand out during the decision process).
Intentional learning favors recall, because the effort to learn entails the deliberate setting up of cues that can aid the later search process, while recognition is the same following intentional and incidental learning, because recognition obviates the need for search.

As intuitively appealing as dual-process theories may be, Craik (based on his notion of "levels of processing") and Tulving (based on the kind of experiments described here) offered a single process theory that focuses on the cue information provided in recall and recognition tests.

Retrieval from memory always begins with some kind of query.
This query contains cue information about the features of an appropriate answer.
This cue information contacts information stored as part of the memory trace (remember Tulving and Bower's analysis of the memory trace, from the lectures on Representation).
Retrieval is a function of the amount of information provided by the retrieval cue.

Thus, the various testing procedures differ in the amount of information provided by the query -- and not just the amount of information, but also the type of information.

In Free Recall the cue information is vague and generalized. The cue provides information about the spatiotemporal context of the event, and the subject must retrieve information about the event itself.
In Cued Recall the cues provide more specific information: they specify not just the context, but also a description of the general features of the event, or related knowledge also stored in memory.
In Recognition the cue is a literal copy of the event itself. Recognition is thus a special case of cued recall, in which the retrieval cue is identical to the target item. Thus, in recognition the cue specifies both the context and the event.

These comparisons among free recall, cued recall, and recognition illustrate the principle of Cue-Dependent Forgetting, announced by Tulving (1974):

The Cue-Dependency Principle

The accessibility of a trace available in memory is a function of the richness and informativeness of the cues used to retrieve that trace from storage.

The more information in the cue, the more likely retrieval will be successful. Elaboration and organization enhance memory, apparently, by increasing the likelihood that the cue will contact relevant information stored in memory.

Involuntary Memory?

The cue-dependency principle implies that retrieval always occurs (when it occurs) in response to environmental cues. But is there any truly spontaneous remembering, in which memories are retrieved even when there are no retrieval cues present in the environment?

Here we need to distinguish between at least three different sources of retrieval cues:

Those contained in formal queries to memory, as in What did you have for dinner last Thursday?
Those contained in objects and events incidentally encountered in the environment, as when someone offers you a chicken stew for dinner, and you remember that you had it for dinner last Thursday.
Those that are self-generated, as when you're trying to think of what to offer someone for dinner, and you try to remember what you gave them last time.

One thing is pretty clear: an event does not have to be consciously attended in order for it to serve as a retrieval cue. But when we remember something, there was most likely a retrieval cue of some kind present somewhere.

Implicit Memory and Recollective Experience

There's more to be said about recall and recognition, particularly about the subjective experience of remembering, but I'm deferring discussion of these topics until after the lectures on Implicit Memory -- which, I hope, that will make that discussion more sensible.

The Relation Between Encoding and Retrieval

We usually consider encoding , storage, and retrieval to be separate phases of memory processing, with two separate factors governing accessibility: processing at the time of encoding and cues at the time of retrieval. But in fact, it is difficult to separate encoding and retrieval -- even by the longest retention interval.

Elaboration and organization involve retrieving information from memory, as the new trace makes contact with pre-existing knowledge, and the subject draws upon pre-existing knowledge in the service of organization.
And the subject must encode the retrieval cue to extract information from it: semantic cues will not be effective if they are not processed semantically, and the subject must often retrieve additional information from memory to supplement the retrieval cue.

Encoding and retrieval processes clearly interact in important ways.

Rich retrieval cues can compensate for poor encoding paradigm -- which is why, in the LoP paradigm, recognition is especially superior to recall for items subject to shallow processing.
And deep encoding can compensate for impoverished retrieval cues -- which is why the recall-recognition difference is diminished for items subject to deep as compared to shallow processing.

The classic mnemonic devices enhance both encoding and retrieval processes.

They foster encoding by fitting the new information into a rich network of prior knowledge.
At the time of retrieval, the devices provide extra retrieval cues.

The principle of cue-dependency makes an important statement about retrieval, which is that retrieval is always initiated by some query to memory -- either occurring in the external environment or generated by the subject. This query supplies cue information that guides retrieval. As a general rule the amount of cue information determines accessibility: the more information available in the cue environment, the more readily will the target trace be accessed. But access is not just a matter of the amount of information. It is also a matter of the type of information.

Encoding Specificity

Consider a famous demonstration by Tulving and Thomson (1973) of the recognition failure of recallable words.

Tulving and Thomson presented subjects with a list of words arranged as cue-target pairs, e.g., ground-COLD. In each case, the cue was a very weak associate of the target, but the subjects were instructed that the cue might help them remember the target on a later memory test.
Then, they tested cued recall by presenting the cue terms, and asking the subjects to remember the associated targets.
This study-test cycle was repeated for another list -- e.g., badge-BUTTON.
And then for yet another, third list -- e.g., glue-CHAIR. The purpose of this repeated procedure was to induce a particular encoding set, in which subjects formed associative links between the cues and the targets.
But after the third study trial, instead of a test of cued recall, the subjects were surprised with a test of free association. They were asked to generate associates to new cue words that had been deliberately selected to be strong associates of the targets -- e.g., table-CHAIR. Naturally, the subjects did generate some targets as associative responses.
The subjects were then asked to examine their list of free associates, and to circle any item that had been on their study list.

The finding was that subjects missed many targets on this recognition test. But when they were presented with the original weak cues, such as glue, the subjects recalled many items that had been previously forgotten on the recognition test.

The recognition failure of recallable words reverses the usual relation between cued recall and recognition. Usually, cued (and free) recall is worse than recognition. But in this case, cued recall is better than recognition. This was true even in an experiment involving 2-alternative forced choice recognition.

The recognition failure of recallable words underscores the interaction between encoding and retrieval processes. How an item is processed at the time of encoding determines not just whether it will be subject to elaborative and organizational processing. It also determines what cues will be effective at the time of retrieval.

Hence, the encoding specificity principle (ESP; Tulving & Thomson, 1973).

The Encoding-Specificity Principle

The accessibility of an event in memory is a function of the overlap between cue information processed at the time of encoding and cue information processed at the time of retrieval.

Put another way, cue information processed at the time of encoding determines what cues will be effective in gaining access to memory at the time of retrieval. All encoded memories remain permanently available in storage. Available memories are accessible to the extent that information supplied by the retrieval cue matches information encoded in the memory trace.

Transfer-Appropriate Processing

Encoding specificity is related to another principle known as transfer-appropriate processing, or TAP (Morris, Bransford, & Franks, 1977):

The accessibility of an event in memory is a function of the overlap

between the processes deployed at the time of encoding and those that are deployed at the time of retrieval.

In TAP, the focus is on the overlap in processes engaged at the time of encoding and retrieval, whereas the ESP focuses on the overlap in the cues being processed.

TAP was famously demonstrated by Morris, Bransford, and Franks (1977) in a variation on the LoP experiment.

Words were presented under one of two orienting tasks:

phonemic (e.g., Does hail rhyme with bail?) or
semantic (e.g., Is hail associated with snow?).

Later, subjects were tested for cued recall, with two types of cues:

phonemic, with a new rhyming word (e.g., pail-____) or
semantic, with a new associate (e.g., sleet-_____).

Morris et al. replicated the levels of processing effect: Overall, there was better memory for items subject to semantic encoding. But more critically, there was a significant interaction:

For items subject to semantic encoding, retrieval was better when given semantic cues.
For items subject to phonemic encoding, retrieval was actually better when given phonemic cues.

This result is not produced by an overlap in cues between study and test, because the cues actually changed, from hail to pail and from hail to sleet. Rather, the effect occurs because the processes -- phonemic or semantic - -remain the same.

A somewhat similar demonstration was produced by Tversky (1973), who asked subjects to study a list of words in anticipation of either a recall or a recognition test. Later, half the subjects in each group were surprised to receive the test of the other kind. Overall, recognition was better than recall. But again, there was a significant interaction:

Subjects anticipating a recall test did better when given a recall test, rather than a recognition test.
Subjects anticipating a recognition test did better when given a recognition test, rather than a recall test.

So, as with Tulving and Thomson (1973), recognition is not always superior to recall. Apparently, subjects expectations about how they were to be tested affected their processing at the time of encoding (there are implications of this experiment for short-answer versus multiple-choice testing in academic context, but we need not go into these here!).

We'll return to TAP later, when we discuss the differences between explicit and implicit memory. But for now, we'll continue to discuss some ramifications of the encoding specificity principle.

State-Dependent Memory and Its Variants

Encoding specificity appears to underlie the phenomenon of state-dependent memory (SDM), where retrieval of a memory depends on the match between the organism's physiological state at the time of encoding, and its physiological state at the time of retrieval.

SDM, in turn, was first observed in an experiment on animal learning by Overton (1964).

In Overton's experiment, rats were trained to run a T-maze, turning right or left at the choice point in order to escape shock. Before learning trials, the rats were drugged with a high dose of barbiturate, and then half the rats were reinforced for turning left, the other half reinforced for turning right. Over 10 training sessions, the animals learned to respond perfectly.

Overton allowed the drug to wash out of the rats' systems, and then retested them in a maze in which escape was possible with either response (for this reason, no further learning was possible during the test trials). The result was that the rats chose the right and left arms of the maze randomly -- it was as if they had forgotten what they had learned.
A control group also learned the maze under the influence of the drug, and went through the washout period, but received a re-administration of the drug before testing. These subjects reliably repeated the response they had previously learned, turning left or right.

Thus, it seemed that the rats' learning was state-dependent, in that correct performance depended on the congruence between the state in which the learning occurred and the state in which the learning was tested.

Because this experiment was conducted before the cognitive revolution took hold in psychology, and because Overton himself had behaviorist leanings, the phenomenon was initially called state-dependent learning -- and, in some quarters, as drug-discrimination learning (because, ostensibly, the drug state served as a discriminative stimulus for the organism's response). But these days, when psychologists do not hesitate to talk of memory in both humans and animals, and when learning is defined as the acquisition of knowledge, the preferred label is state-dependent memory.

SDM has also been observed in humans. In the typical SDM design, words are presented for study, and then tested for free recall, while manipulating the subject's internal physiological state by administering psychotropic drugs (like barbiturate, amphetamines, marijuana, alcohol, nicotine, and even caffeine) that act directly on the central nervous system. The experimenter then varies the congruence of the drug state between encoding and retrieval.

In the congruent conditions, the subject is in the same drug state during both phases.
In the incongruent conditions, the subject is in a different state during each phase.

Typically, memory is best when the subject is drug-free in both conditions. This is because most psychotropic drugs impair memory, either by impairing encoding operations or by impairing retrieval operations, or both. We don't have to get into the physiology here: the short version of the story is that psychoactive drugs generally impair the deployment of cognitive resources. But if the subject receives a drug, memory is better if he receives the drug in both phases, as opposed to just one.

Put another way, if we plot the probability of recall against encoding-retrieval congruence, recall is best if it occurs in the same state, as opposed to a different state, present at the time of encoding. That's the basic SDM phenomenon.

Here's an actual example, from a study by Swanson and Kinsbourne (1976) on the effects of Ritalin on leaning and memory. Ritalin is an amphetamine-like stimulant which usually impairs learning. But it has well-known "paradoxical" effects on subjects with attention deficit-hyperactivity disorder (ADHD), improving their performance (apparently, their ability to pay attention) in many domains.

In their experiment, S&K asked hyperactive and control children to perform a variant on paired-associate learning called a zoo-location task, which paired 48 animal names with one of 4 familiar cities, as in elephant-Vancouver (the subjects were Canadian).

On Day 1, half the subjects in each group learned half the list while drugged, and the other half learned un-drugged.
On Day 2, the subjects in each group relearned the list while drugged or un-drugged.

In terms of errors in initial learning, it is clear that Ritalin impaired learning in the control children (more errors), as would be expected (because it's a psychoactive drug), but had a paradoxical effect on the hyperactive children, actually improving their learning performance.

On Day 2, those hyperactive children learned fastest whose drug states matched their state on Day 1: the congruence between encoding and retrieval states apparently enhanced memory. There were similar, though weaker, results for the normal subjects.

By now, SDM has been shown, to varying degrees, by a lot of centrally acting drugs -- not just anesthetics like barbiturate, but also anti-anxiety and antidepressant agents, narcotics, hallucinogens, and even nicotine (see the review by Eric Eich). The SDM effects of caffeine are dubious -- perhaps because, between coffee, Coke, and Pepsi, not to mention Mountain Dew, Red Bull, and other "energy drinks", the caffeine administered in the laboratory is like a drop in the ocean. There are no effects, apparently, of aspirin or lithium (a common treatment for bipolar disorder).

SDM effects are relatively weak in humans -- partly because the doses of psychoactive drugs that can be ethically given to humans are weaker than those that can be administered to animals; and partly because subjects can generate their own retrieval cues, which may overshadow the cues provided by the drugs alone. (In Overton's experiments, the dose of barbiturate employed was huge, just barely sub-lethal, and it is amazing that, stumbling down the maze, that the rats learned anything at all.)
SDM effects are also asymmetrical. Generally, there is little transfer from the drug condition to the control condition, but more transfer from the control condition to the drug condition.
As noted earlier, there is usually a main effect of the drug. Centrally acting drugs nearly always impair learning and memory -- which is why it's probably not a good idea for people who don't suffer from ADHD to take amphetamines to "boost" their "brain-power".

The exception, of course, is Ritalin in patients with ADHD.
We don't know whether there are similar paradoxical effects in other groups -- for example, whether anti-anxiety agents improve cognitive performance in anxious patients, or whether antidepressants improve cognitive performance in depressed patients.

As Eich noted, SDM is cue-dependent.

In humans, SDM is most reliably obtained with tests of free recall.

The principal exception is the Swanson & Kinsbourne (1976) study, in which the paired-associate procedure constitutes a combination of cued recall and savings in relearning.
Otherwise, SDM tends not to occur with cued recall or recognition tests.

The cue-dependency of SDM reminds us that there are two kinds of cues involved in the phenomenon:

Nominal cues, which are explicitly provided at the time of retrieval; and
State cues, background cues provided by the drug state.

Apparently, the nominal cues can overshadow the state cues. In free recall, the nominal cues are relatively impoverished, but in cued recall and recognition they are much stronger -- strong enough, apparently, to overshadow the relatively weak state cues provided by the relatively weak drug states used in humans. This might not necessarily be the case with really strong drug states -- but again, ethical considerations preclude this kind of study.

SDM is not limited to drug states.

Eich and Metcalfe (1989) induced happy or sad emotional states by having their subjects listen to music, and then had them study a list of words. In a variant on the LoP paradigm, they read half the words from a list (e.g., vanilla); the remaining words were generated from cues (e.g., Milkshake flavors: Chocolate and _____). Typically, words that are generated by the subjects themselves are remembered better than those that are merely read -- a phenomenon called the generation effect, which is generally interpreted in terms of the elaborate processing induced by the task of generating the target words.

Anyway, on a later recall, Eich and Metcalfe induced the same or different mood in their subjects, again by listening to music.

They observed the generation effect, such that overall, memory was best for self-generated items.
More important, they observed an interaction, such that memory was best when the subjects' mood at the time of retrieval matched their mood at the time of encoding. This is known as mood-dependent memory.
Notice, as well, that there was a main effect of mood.

In general, sadness impairs memory, at both encoding and retrieval --perhaps because depression consumes cognitive resources, or because depression lowers the subject's motivation.

Drug states and emotional states are both aspects of internal context -- the subject's internal physiological state, or the subject's internal mental state. But we can get similar effects by manipulating external context.

For example, Abernathy (1940) noted that college examination scores were higher when the exam was given in the student's usual classroom, with the usual instructor serving as proctor.

Some studies employ a more radical manipulation of the environment at encoding and retrieval. For example, Godden and Baddeley (1975) found evidence of environment-dependent memory (EDM) in SCUBA divers who studied a list either on the beach or 15 feet underwater, and who received a free recall test either on the shore or under water. Memory was best when study and test took place in the same environment.

So far as we know, EDM follows the same rules as drug state-dependent memory (see reviews by Smith, Glenberg, and Bjork), and papers by Fernandez and Glenberg).

There are "dose" effects, such that the bigger the difference between encoding and retrieval environment, the bigger the effect.
There is a cue-dependency, such that EDM is obtained only with free recall tests (there is, apparently, no environment-dependency on multiple-choice exams, so students in the introductory course, and other large courses, can rest assured that they are getting the grades they deserve).

Interestingly, EDM effects can also be abolished by the subject. Changing the physical environment can induce EDM, but EDM can also be reversed if the subject is simply asked to imagine being in the original environment. Apparently, it's the mental representation of the environment that matters.

State-dependency, emotion-dependency, and environment-dependency are all aspects of a general effect of context-dependent memory. Apparently, information about the context in which an event occurred is encoded as part of the memory trace.

Usually we think of context in spatiotemporal terms -- that is, the time and place that an event took place.
But context can include other features of the external context, such as a description of the physical environment.
And context can also include features of the internal environment, including the subject's physiological state, emotional state -- and, perhaps motivational state as well. Do subjects who study hungry remember better when they are hungry?

All of this contextual information appears to be processed, to at least some extent, along with the nominal target items, and it is also processed, to at least some extent, along with the retrieval cues.

In congruent conditions, there is more overlap between encoding and retrieval cues, so it is more likely that the retrieval cues will find a match with encoded trace information.
In incongruent conditions, cue information doesn't match trace information as well, so that the retrieval process may miss relevant traces available in memory, and the subject will fail to gain access to available memories.

Schematic Processing

The encoding-specificity principle (and the related principle of transfer-appropriate processing) illustrate the general point that encoding and retrieval factors work together to determine the accessibility of memory.

Encoding sets up the best route to successful retrieval.
Retrieval recapitulates the processes deployed at the time of encoding.

We can also see the joint effects of encoding and retrieval processes in another phenomenon, known as the schematic effects on memory.

As noted in the lectures on Representation, the schema concept in psychology has its origins in the work of F.C. Bartlett, and his research on memory for stories.

To make a long story short, a schema is an organized knowledge structure that represents our knowledge and expectations about the world -- something like a concept, or more accurately a set of concepts.
Cognitive schemata serve to organize perception and memory, such that we perceive and remember events in terms of the schemata that are active at the time.
The effects of schemata reflect what Bartlett called "effort after meaning". The subject is not merely a passive recipient of sensory information, or a passive repository of memory. Rather, in perceiving and remembering the person is actively engaged in intelligent problem-solving: using all the resources at his command to figure out what is happening, or what happened in the past.

Interest in cognitive schemata disappeared during the heyday of associationism and behaviorism,

Behaviorism didn't like references to internal mental structures, and preferred to analyze behavior in terms of stimulus-response connections.
Associationism preferred to analyze internal states as elementary associations, and had little interest in larger cognitive structures.

But more recently, interest in schemata was revived within cognitive and (cognitive) social psychology (Rumelhart & Norman, 1975; Taylor & Crocker, 1981). According to the organizational principle, for example, concepts, stories, and the like provide a framework for encoding and retrieval. And interest in schemata increased as cognitive psychology began to analyze memory for more lifelike stimulus materials, such as stories and social behavior.

The schema concept was introduced into cognitive psychology by Bartlett (1932), who adapted it from earlier theoretical work of Sir Henry Head on bodily posture (Head & Holmes, 1911). Head noted that an organism receives a considerable amount of sensory information about the position and activity of its own body -- Sherrington (1904) had referred to these sensations as proprioception. This has to be integrated with some ongoing sense of what the body's posture is currently -- a sense that has to be continuously modified in light of incoming proprioceptive feedback. This evolving mental image is the body schema. Bartlett took this basic concept of the body schema and applied it to cognition. In his terms, a schema is the cognitive background against which perception is constructed, and memory is reconstructed, and is itself modified by those percepts and memories. A schema is an organized knowledge structure containing generic knowledge, beliefs, and expectations. It organizes perception and memory -- by which Bartlett means perceiving and remembering. As such, schemata provide the cognitive basis for the "effort after meaning" that is central to Bartlett's view of mental life.

Of course, a very similar concept of schema is central to Piaget's theory of cognitive development, as introduced in The Child's Conception of the World (1926). For Piaget, a schema is a cognitive background against which assimilation and accommodation take place.

Bartlett's schema concept was roundly criticized by fellow British psychologists (Oldfield & Zangwill, 1942), some of whom found Bartlett's use of the term incomprehensible. And in the United States, his work was largely ignored by the functional behaviorists who dominated American academic psychology. The idea of the schema was revived in experimental psychology by Neisser (1967), who made the concept central to his textbook on cognitive psychology, and in clinical psychology by Beck (1967), who made the concept central to his cognitive theory of depression and its treatment.

The relationship between Bartlett's and Piaget's use of the schema concept is problematic, and is a problem for the historians of psychology to resolve. Piaget published before Bartlett, in 1926, but Bartlett does not cite him -- his exposition is derived exclusively from Head. In somewhat the same way, Neisser and Beck do not acknowledge each other, even though they were writing their books at roughly the same time (the mid-1960s) and in the same place (the University of Pennsylvania, where Neisser was on sabbatical from Cornell in Martin Orne's Unit for Experimental Psychiatry). However, both Neisser and Beck were aware of Schachtel's paper which applied Piaget's theory to infantile and childhood amnesia, and Neisser cites Schachtel in his book.

There ensued, especially in the 1970s and 1980s, a full-scale "Bartlett revival" in the study of memory, as exemplified by the work of Gordon Bower and his students on story memory (for a review, see Brewer, 2000). Moreover, the schema concept was embraced by the emerging cognitive perspective in social psychology.

According to Taylor and Crocker (1981), schemata had a number of cognitive functions:

Schemata lend structure to experience.
They determine what will be encoded and what will be retrieved.
They speed up processing time.
They fill in missing information, permitting the perceiver, in Bruner's felicitous phrase, to "go beyond the information given" by the stimulus.
In a similar manner, they provide a basis for problem-solving, evaluation, and planning.

Perhaps the most thorough explication of the schema concept was its application to artificial intelligence by Rumelhart and Ortony (1976), following earlier work by Minsky (1975). They begin by assuming that a schema is a special type of data structure that captures conceptual knowledge.

They argued that every schema has a slot structure consisting of a fixed core and variable slots. To take an example from Anderson's Cognitive Psychology and Its Implications (5e, 2000), consider the schema for house.

The core is that a house is a a type of building used for housing humans.
Slots contain information about variable features of the core, such as how many rooms there are (and what their types are), whether it is made of brick or wood, whether it is small or large.
The schema itself contains default values for each of the slots -- indicating, for example, that the typical house has six rooms, including 2-3 bedrooms and 2 bathrooms, and occupies about 1800 square feet.

Every schema is embedded in a structure of superordinate and subordinate schemata (or, if you prefer, schemas).
Schemata are generalized at an intermediate level of abstraction: more specific than a "building" or even a "dwelling", but more abstract than (3BR 2-1/2Ba on .25 acres off a cul-de-sac").
Schemata aid in comprehension, mediating both "bottom-up" (accommodation) and "top-down" (assimilation) processes. Schemata tell the perceiver what to expect, and are modified by what the perceiver finds.

Schema-Relevance and Schema-Congruence

what argued that schemata determine what will be encoded in, and retrieved from, memory. For Bartlett, encoding favored schema-congruent information -- information that fit with the person's expectations, and that was easily assimilated into currently active structures. And he suggested that schema-incongruent information might be ignored, and thus not encoded at all -- or else, distorted so as to be assimilated into prevailing schemata. Bartlett also argued that schemata provided the basis for retrieval (or, as he preferred to call it, reconstruction). Initially, he argued, the person remembers the event in very general terms, and then invokes a schema which guides the rest of the retrieval/reconstruction process.

When the schema concept was revived, it was naturally revived in the context of memory -- in particular, examining the relations between particular events and generic schemata, viewed as a kind of semantic knowledge, and their effect on subsequent episodic memory. The result was a bunch of conflicting findings: pretty much everyone found that schema-congruent events were well remembered, but the fate of other kinds of items was unclear. And, in fact, Bartlett's own evidence was very ambiguous on this point -- partly because he was unclear about what a schema was, and also due to his aversion to quantitative analysis.

The situation was clarified by a series of papers by Hastie (1980, 1981), who began by noting that there were three different relationships that could obtain between general schematic knowledge and particular events:

Schema-congruent events are highly likely, given a particular schema.

In mathematical terms, p(event | schema) > p(event | no schema)

Schema-incongruent events are highly unlikely given that schema.

p(event | schema) < p(event | no schema)

Schema-irrelevant events are simply not predicted by the schema one way or another.

p(event | schema) = p(event | no schema)

The problem with earlier research on schemata and memory was that it did not make a clear distinction between schema-incongruent and schema-irrelevant events. And it turns out that this distinction actually makes a difference.

Hastie and Kumar (1979) studied the schematic effects on memory in the context of person memory, a topic in social cognition that is concerned with memory for the attributes and behaviors of other people. In this context, knowledge of a person's general attributes comprises the schema for that person; and knowledge of his specific behaviors constitutes episodic memory.

In their experiments, subjects wee presented with information about a particular person in the form of a trait ensemble, or a list of personality traits ostensibly characteristic of that person.

So, for example, a particular person, Judy might be described as intelligent, intellectual, artistically sensitive, refined, imaginative, and witty.
The ensemble was intended to induce a unitary, general impression of the target person's personality -- his or her traits, attitudes, and interests -- in this case, as intelligent and sophisticated.

Then, the subjects studied a list of actions ostensibly performed by the target person.

Some of these were schema-congruent, in that they had a relatively high probability of occurring, given the description of the target, e.g.,

Judy won the chess tournament.
Judy attended the symphony concert.

Others were schema-incongruent, in that they had a relatively low probability of occurring, given the description of the target --e.g.,

Judy made the same mistake three times.
Judy was confused by the TV show.

And finally, others were schema-irrelevant, or schema-neutral, in that they did not relate to the schema one way or another -- e.g.,

Judy ordered a sandwich for lunch.
Judy took the elevator to the 3rd floor.

Later, H&K tested free recall of the behaviors, and plotted the probability of recall against schema-congruence.

In Experiment 1, with 4 events in each category, H&K obtained a U-shaped function, such that memory was worst for schema-irrelevant items, and better for schema-incongruent items than for schema-congruent items.
Examining the role of input serial position, they replicated the U-shaped function at 3 of four serial positions.

Of course, it would be unusual to have equal numbers of schema-congruent and schema-incongruent events. Almost by definition, schema-incongruent events have to be relatively rare -- otherwise, you'd have a very different schema for that person. Schema-incongruent items, precisely because they violate our expectations, should be relatively infrequent. Accordingly, in a subsequent experiment H&K constructed alternative lists of 12 behaviors:

12 congruent, 4 neutral, and 0 incongruent.
11 congruent, 4 neutral, and 1 incongruent.
9 congruent, 4 neutral, and 3 incongruent.
6 congruent, 4 neutral, and 6 incongruent.

Memory always favored schema-congruent over schema-neutral behavior, and schema-incongruent over schema-congruent behavior.

Hastie (1984) conducted a further experiment in which the trait ensemble was followed by a list of schema-congruent and schema-incongruent items. Recall testing yielded the schematic processing effect, as expected. However, in a second experiment Hastie asked subjects to perform a sentence-continuation task: after each item, they were supposed to continue it with either an explanation of the event, and elaboration of the event, or the sequel to the event. On a later recall test, items (whether schema-congruent or schema-incongruent) in the explanation condition were recalled better than those in the elaboration or sequel condition. So, it's not schema-incongruency per se that yields better memory: it's the explanatory activity that schema-incongruency instigates.

These experiments illustrate the schematic processing principle:

The Schematic Processing Principle

Memory for an event is a function of the relationship that event and pre-existing schematic knowledge, expectations, and beliefs.

Schematic processing actually reflects two different processes affecting schema-congruent and schema-incongruent information.

Schema-congruent events are predicted by the schema, and fit right in: They are easily encoded with respect to the prevailing conceptual framework.
Schema-incongruent events are not predicted Their occurrence must be explained, and this explanatory activity promotes extra processing at the time of encoding, and yields a more deeply encoded, more elaborate, memory trace.

Thus, in a sense:

Schema-congruent events get the benefit of organizational activity, because the schema provides an organizational framework at the time of retrieval.
Schema-incongruent events get the benefit of elaborative activity at the time of encoding.
Schema-irrelevant events get no benefits at all: They are not unexpected, so they don't get extra elaboration; and there is no general schema to provide relevant cues at retrieval.

The principle of schematic processing brings us full circle -- reviving, in the analysis of retrieval, the earliest discussion of the importance of elaboration and organization at the time of encoding.

Schematic Effects in a Network Model of Memory

Srull (1981) offered a somewhat different explanation than Hastie for schema-dependency, within the framework of a generic associative-network model of memory. He proposed that nodes representing individual episodes are linked to a node representing the person, in the usual way. Then, connections among nodes are produced by virtue of processing at the time of encoding -- such as explaining schema-incongruent items in light of the schema. However, nodes representing schema-incongruent items are associatively linked both to each other and to nodes representing schema-congruent items as well.

Testing recall, Srull obtained the usual schema-dependency effect. Schema-relevant items were recalled better than schema-irrelevant items, and schema-incongruent items were recalled better than schema-congruent items.

Then, Srull employed a sentence-verification procedure, not unlike that which had been used by Anderson & Hastie (1974), to examine priming effects on recognition memory. Srull compared response latencies to verify schema-congruent, incongruent, and -irrelevant items, depending on the immediately preceding item. Compared to a baseline provided by schema-irrelevant items, schema-congruent items primed responses to schema-incongruent items, while schema-incongruent items primed both schema-congruent and schema-incongruent items; schema-irrelevant items didn't prime anything. These results are consistent with Srull's hypothesis, that schema-incongruent items are linked to each other and to schema-congruent items, but that schema-congruent items are not directly linked to each other.

Reconstruction

So far, the methods used to study memory, and the basic principles derived from those methods, wouldn't have surprised Ebbinghaus. If he had lived to 1980, he would have been impressed by the progress made beyond the Law of Repetition and the Principle of Time-Dependency, and he certainly would have been impressed by our advances in understanding the biological bases of memory, but he would view this progress as natural, and not embodying anything like a paradigm shift. This is because most of the work so far has been based on a more or less sophisticated version of the library metaphor, with memory traces, representing events, being encoded, stored, and retrieved much like books on a library shelf.

Lying behind the library metaphor is a particular view of the memory trace, as something that has an existence independent of the person doing the remembering. The trace exists "in memory", and must be found in order for remembering to proceed.

Bartlett and Remembering

But an entirely different approach to memory was introduced by Frederick C. Bartlett (1932), a pioneering British psychologist (British, Canadian, and Australian psychologists like to trace their heritage to Bartlett, much like American psychologists like to trace their heritage to William James).

Some hint of the difference can be seen in the titles of Ebbinghaus's and Bartlett's books:

Ebbinghaus wrote Uber das Gedachtniss, translated as On Memory, with memory as a noun -- a representation of a thing.
Bartlett wrote Remembering: A Study in Experimental and Social Psychology, with remembering as a verb -- a representation of action.

And the difference can also be seen in their approach:

Ebbinghaus, practicing the elementarism promoted by the psychophysicists and the structuralists, sought to strip memory from its personal and social context.

True to his psychophysical heritage, Ebbinghaus attempted to control every independent variable and measure every dependent variable. He was the experimenter par excellence.

Bartlett's view was that this was impossible, and that the attempt so distorted the phenomenon that laboratory research based on Ebbinghaus's approach bore essentially no relation to memory as encountered in the real world outside the laboratory.

Bartlett, working under the influence of anthropology, focused on observation, and considered control and quantification to be not so important. He was, truth to be told, a terrible experimenter.

Bartlett's work arouse out of a basic objection to Ebbinghaus's approach. Ebbinghaus focused on rote memorization of meaningless materials -- he invented the nonsense syllable, and his experimental procedures focused on serial learning. His underlying view of memory was as reproductive -- that good memory preserves accurate representation of the past. And by focusing on serial learning, Ebbinghaus forced his subjects to make arbitrary associations among stimuli.

But in Bartlett's view, Ebbinghaus's procedures, and findings, were misleading, and did not represent memory as it operated in the real world. In the real world, Bartlett argued, memory is reconstructive, not reproductive. People don't retrieve memories of past experiences; rather, they reconstruct memories.

The memory trace, such as it is, does not contain a complete mental representation of some past event. Rather, trace information is vague, fragmentary, and ambiguous.
Trace information must be combined with other information in the process of remembering.

Information contained in retrieval cues: the query to memory contains information about the memory being sought.
Inferences based on world knowledge: abstract knowledge about the to-be-remembered event.

"The War of the Ghosts"

Bartlett began his treatise on Remembering: A Study in Experimental and Social Psychology with a critique of the verbal-learning paradigm invented by Ebbinghaus for Uber das Gedachtniss. In fact, he began it with a critique of 19th-century psychophysics, which he then extended to Ebbinghaus. For Bartlett, the whole enterprise was too sterile, because the stimulus materials, and what subjects were asked to do with them, were devoid of meaning -- effectively denying subjects the effort after meaning which he considered essential to understanding mental function. Ebbinghaus, following Fechner and the other 19th-century psychophysicists, invented the nonsense syllable precisely to maintain tight control of stimulus conditions; and the nonsense syllable, by its very nature, was intended to frustrate subjects efforts after meaning, and force them to form merely rote associations between one meaningless CVC and another. The whole thing was wrongheaded. As Bartlett famously put it -- and every psychologist should keep this inscribed in a wallet-sized card:

"The psychologist, of all people, must not stand in awe of the stimulus."

In a foreshadowing of Martin Orne's critique of ecological validity, Bartlett argued that remembering was not adequately represented by the rote memorization of unrelated meaningless items -- what the psychologists of the Bartlett Revival disparaged as "grocery lists". Remembering was less like rote recitation and more like the telling of stories. Accordingly, in his experiments Bartlett told his subjects unfamiliar stories -- a favorite was "The War of the Ghosts", a folktale collected from Native Americans in the Northwest by the pioneering anthropologist Franz Boas, which begins "One night two young men from Egulac...". He read the story out loud twice. Then, after a suitable retention interval, Bartlett then asked his subjects to tell the story themselves. For this purpose, he employed two somewhat different methods.

In the method of repeated reproduction, individual subjects were asked to repeat the story on several different occasions.
In the method of serial reproduction, one subject was asked to tell the story to a second subject, who in turn told the story to a third subject, who then told the story to a fourth subject.... -- in a laboratory version of the parlor game of "telephone".

Examining how the stories changed over repetitions, Bartlett naturally observed progressive forgetting -- especially of minor details, but not so much of the gist of the story. But he also observed other changes that were not easily subsumed under the heading of time-dependency.

Errors of commission as well as of omission (otherwise known as forgetting): Subjects often imported new details into their stories, which were not contained in the original.
Rationalization, new material intended to explain or justify the events of the story.
Transformations of detail, to make the story more familiar -- for example, replacing the Indians' "canoe" with "boat".
Transformations of order, to give the story more of a conventional narrative structure.

Based on results such as these, Bartlett argued that remembering was reconstructive, not reproductive, in nature.

Remembering begins with a question about the past.
The subject's initial response is to retrieve the dominant details of the event in question. This trace information is typically vague, fragmentary, and ambiguous.
The subject also retrieves a general attitude toward the story.
Remembering proceeds on the basis of schema-based inferences about the story, based on the knowledge, beliefs, expectations, and attitudes contained in whatever schema is brought to bear on the task of remembering.
The result is a more or less coherent story about the past that may or may not be historically accurate.

Which leads us to the last principle governing memory:

The Reconstruction Principle

Memory reflects a blend of information retrieved from specific traces with knowledge, expectations, and beliefs derived from other sources.

The Reconstruction Principle qualifies the Library Metaphor so frequently invoked (including by myself) as a framework for understanding memory.

Remembering is not like finding a book on a shelf and reading its contents.
Rather, remembering is more like writing a book based on fragmentary notes.

The "Bartlett Revival"

Bartlett published his book in 1932, and then pretty much left further research to others. He wasn't a great methodologist to begin with, and reading through the "experimental" portions of Remembering, you get the sense that like James, he understood that research was crucial but his heart just wasn't in it. (Though he did produce a string of distinguished memory researchers, including Graham Reed, author of The Psychology of Anomalous Experience, which has extended treatments of such phenomena as deja vu.) And, as noted, his work was ignored for a long time: his British colleagues didn't understand it, and American psychologists were too infatuated with behaviorism. But beginning with Neisser's references to schema theory in Cognitive Psychology (1967), cognitive psychology, and especially memory research, underwent a sort of "Bartlett Revival" in which investigators began to study memory for things other than lists of nonsense syllables, words, and pictures, and in which theorists began to employ the schema concept without embarrassment.

There is an interesting story here. Bartlett himself rejected the verbal-learning procedure initiated by Ebbinghaus as too devoid of meaning. And in 1978, Neisser himself castigated those who employed the verbal-learning paradigm for sacrificing ecological validity to the appearance of methodological rigor:

If X is an interesting or socially significant aspect of memory, then psychologists have hardly ever studied X.

Both Bartlett and Neisser had a point, as anyone who has gotten through the lectures on Associationism and Interference Theory can attest. But it also turned out that they were wrong. Once researchers grasped what Bartlett had been up to, and once the Cognitive Revolution was thoroughly ingrained in psychologists' mind and behavior, it was pretty easy to come up with ways to explore reconstructive processes with the constraints of the verbal-learning paradigm.

The Semantic Integration Effect

Among the earliest such efforts was research on the semantic integration effect by Bransford and Franks (1971, 1972).

	They asked their subjects to study sentences of the form The girl broke the window on the porch, The girl lives next door, and The window was large.
	A subsequent recognition test included lures like The girl who lives next door broke the large window on the porch, which contained propositions derived from the studied targets.
	The principal finding was a high level of false recognition -- especially for lures which contained several ideas from the studied items.

This is reconstruction in almost pure form. The subjects have "constructed" a memory based on fragmentary material, woven into a sort of story.

Point of View Effects

An even more explicitly "Bartlettian" approach is exemplified by research on story memory -- that is, memory for narratives, instead of the usual word lists.

One line of research focused on point of view effects -- that is, the effect on memory of taking the point of view when reading a story. For example, Bower and his associates asked subjects to read a story about two boys playing hooky from school, which described their activities around the house. The subjects were asked to read the story from the perspective of either a home-buyer or a house-burglar. After recalling the story, they were asked to shift to the other perspective. In this condition, the subjects recalled new details that were important to the second perspective, but not the first.

The Post-Event Misinformation Effect

A famous example of reconstructive memory was provided by Loftus and her colleagues, in a series of studies of eyewitness memory. The study employed a laboratory model of eyewitness memory, and thus has all the appearances of ecological validity, but deep down it's just a variant on the verbal-learning paradigm. The stimulus materials are pictures or film clips, not words, and the stimuli are ordered into a narrative instead of randomly, but it's a long ways from "The War of the Ghosts" (this is not a criticism of Loftus, but rather an illustration of the point I made earlier: reconstructive processes can be studied using standard verbal-learning paradigms).

In one study, Loftus & Palmer (1974) showed subjects a film of a staged automobile accident. Later, the subjects were "interrogated" with a set of questions, including "How fast were the cars going when they _____ with each other?". For different groups of subjects, the blank was filled in with one of the following verbs: contacted, hit, bumped, collided, or smashed.
	Then the subjects were asked to estimate the speed at which the cars were traveling. These estimates varied systematically, depending on the verb inserted into the blank.
	One week later, the subjects were asked if they had seen any broken glass at the accident scene (in fact, they had not). Again, the percentage of subjects reporting broken glass varied with the verb inserted into the leading question.
Subjects who gave relatively high speed estimates in the initial interrogation were more likely to falsely report having seen broken glass.

In another study, subjects viewed a film clip of a car colliding with a baby carriage (Loftus, 1975). During the interrogation, subjects received a number of direct questions, such as "Did you see a school bus in the film?", while other subjects received indirect questions, such as "Did you see the children getting on the school bus?". The direct questions explicitly assumed the presence of a target object, such as a school bus, while the indirect questions merely implied that the object was present. A control group received no questions at all.
	One week later, all subjects were interrogated with the direct questions. The control subjects were the most accurate in their responses; the subjects who had initially received the direct questions were less accurate, but those who had received the indirect questions were even less accurate.

	In the final, and perhaps best-known of these studies, Loftus and her colleagues showed her subjects a slideshow depicting a staged traffic accident in which an automobile struck a pedestrian (Loftus, Miller, & Burns, 1978). In the slideshow, some subjects saw a red Datsun at a corner marked by a "Yield" sign; others saw a "Stop" sign instead.
	Later, they were interrogated about what they saw. Subjects in the control group received a non-leading question, which referred to slides that the subjects actually saw: Did you see another car pass the red Datsun while it was stopped at the Yield sign? Subjects in the experimental condition received a leading question which assumed something that they had not actually seen. Did you see another car pass the red Datsun when it was stopped at the Stop sign?
	On a subsequent recognition test, they were asked to choose between a slide depicting a red Datsun at a Yield sign and one at a Stop sign. And yes, of course, everything was counterbalanced. Loftus was, after all, a product of the verbal-learning tradition, and she knows how to do these experiments! So, Some subjects saw a Yield sign, but received the misleading question about a Stop sign. Some subjects saw a Stop sign, but received a misleading question referring to a Yield sign.
	The result was that recognition of the correct slide was significantly reduced among subjects who received the misleading interrogation.

Again, this is a good example of reconstruction in memory. The subjects saw what they saw, but they might not even have noticed the traffic sign. But the later interrogation assumed the existence of a Stop or Yield sign, and this knowledge was incorporated into their memories. Some of the memory came from trace information, others from the query itself -- and then it was all put together.

These experiments illustrate the post-event misinformation effect. Memory is not "pure", and leading questions can influence eyewitness reports. Apparently, misinformation gleaned from leading questions can be incorporated into an observer's memory for an event.

There are lots more experiments along these lines.

Loftus herself became involved in the controversy over recovered memories of childhood sexual abuse, suggesting that many of such memories may actually be false, a product of misinformation from various sources (including overeager therapists).
Jonathan Schooler, one of Loftus's students, has performed a number of experiments on what he has called verbal overshadowing, in which verbal descriptions can distort memory for various sensory-perceptual experiences. Schooler has also become famous for documenting a decline effect in his own research, in which the strength of the verbal overshadowing effect has diminished over subsequent years. Schooler is not alone in this, but he gets credit for, bravely, drawing attention to the effect as it appears in his own research.

The post-event misinformation effect is important both practically (for its legal implications) and theoretically (for what it can tell us about reconstructive processes), but there are some constraints on the effect. You can't just create a memory for anything, out of whole cloth, by means of leading questions and other suggestive techniques.

The leading questions must be relatively subtle. If they're obviously leading, subjects are likely to notice this and discount their influence.
The leading questions must also be plausible. Of course, this raises the question of what "plausibility" means. In the big scheme of things, it might not be plausible that a child was sexually molested by her schoolteacher (these incidents are, thankfully, relatively rare). But when a patient comes to a therapist for help with anxiety or some other problem, and the therapist (with all the sapiential authority that he or she has by virtue of his or her professional qualifications) tells the patient that her problems may be caused by repressed memories of childhood sexual abuse, then the notion becomes a lot more plausible.
Warning subjects about the possibility of leading questions generally prevent the misinformation effect -- but only if they are given before the misinformation is delivered. If the warning is presented afterward, apparently, the misinformation has already done its work.

What's really going on in the post-event misinformation effect? In her earlier work, Loftus (1975) proposed a construction hypothesis. According to this avowedly Bartlettian approach, perception leaves fragmentary traces in memory, and the leading questions shape reconstruction at the time of retrieval. Later, however, she entertained a substitution hypothesis (Loftus & Loftus, 1980), according to which later events (like receiving the misinformation) overwrite the original trace (derived from perception). The substitution hypothesis assumes that something like retroactive interference or "unlearning" causes a permanent loss of the earlier trace from memory, leaving only the product of the misinformation. The substitution hypothesis is, of course, a direct contradiction of the construction hypothesis. Note, too, that the substitution hypothesis is generally inconsistent with the view that, once encoded, memories remain permanently available in storage, and the mechanism for time-dependent forgetting is interference at the time of retrieval.

Of course, there are other alternatives. One, known as the source hypothesis, states that both memories are available in storage -- the original product of perception, and the later product of leading questions exist side by side. The post-event misinformation effect occurs because the subject is confused as to the source of the memories. The source hypothesis is inspired by findings of source amnesia in amnesic patients and normal subjects, which shows that subjects can remember information that they have acquired through learning, but forget the details of the learning experience.

A variant on the source hypothesis, and also consistent with the construction hypothesis, is the bias hypothesis proposed by McCloskey and Zaragoza (1985). They propose that the post-event misinformation effect occurs when memory for the original event is relatively poor. In their view, the post-event misinformation has no effect on the original memory. Rather, the misinformation creates a separate memory. In the absence of a strong original trace, retrieval and reconstruction is biased toward the trace representing the more recent misinformation.

To understand the bias hypothesis better, consider the following hypothetical classification of subjects:

Those who remember nothing, and guess randomly on the critical test.
Those who remember only the original event, and so can't show the misinformation effect.
Those who remember only the misinformation, and who must show the misinformation effect.
Those who remember both the event and the misinformation, but who are biased toward the more recent information.

In order to test the bias hypothesis, McCloskey and Zaragoza constructed a modified recognition test.

Subjects viewed a slideshow depicting a man in a workshop; at one point he holds a hammer.
Then the subjects are presented with a verbal narrative, in one of two conditions:

One refers to the man holding a wrench.
The other makes no reference to any tool.

On a recognition test, subjects are given a choice between the man depicted holding a hammer or a screwdriver.

Yes, you read that right. Here's the idea (and it's brilliant):

According to the substitution hypothesis, the misled subjects should be less accurate than the controls, because their memory for the wrench will overwrite their memory for the hammer.
According to the bias hypothesis, the misled subjects should perform the same as the controls, because neither alternative provided on the recognition test is consistent with the misinformation

In the original version of the experiment, the misled subjects were less accurate than the controls. This essentially replicates Loftus's methods and results. But in the modified version of the experiment, the two groups showed equivalent levels of accuracy, consistent with the bias hypothesis. Later studies also yielded results consistent with the bias hypothesis -- or, at least, inconsistent with the substitution hypothesis.

So, at least for now, thee post-event misinformation effect remains an example of reconstructive processes in memory. Memory does not involve just the retrieval of a trace from storage. Memory involves judgment and decision-making. Remembering is a judgmental process that takes account of all available information, not just trace information. The misinformation effect shows how, in remembering as in perceiving, the person "goes beyond the information given".

The Associative Memory Illusion

The Bartlett Revival in memory culminated in a large number of studies of what is now known as the Associative Memory Illusion -- also known as the "Deese Roediger McDermott Effect", or simply the "DRM Effect", because it was initially discovered by Deese (1959) and rediscovered and studied thoroughly by Roediger and McDermott (1985). The vast literature on the AMI is reviewed by Gallo (2000, 2006) -- who, because he was a student of Roediger's, calls it the "DRM effect".

	The AMI begins with lists of common free associates to familiar words. For example, given the word needle, the word that comes to mind most frequently is thread, then pin, etc. There are norms for this -- based on data collected decades ago, but good enough to serve the purpose. Subjects then study lists consisting of these associates to the target word (there are several such lists, each targeting a different word).
	Then they are given a recognition test consisting of three kinds of items: Old items (previously studied, like thread and pin). Critical lures (related to the old items, like needle). New items (unrelated to the old items, like sleep). The result is a high rate of false recognition critical lures. That is, subjects falsely remember having studied needle.
	Here's how the effect is supposed to work. Recall that the old items, like thread and pin, were generated as "forward" associations from needle. The AMI is induced during the encoding phase, when items like thread and pin prime needle through "backward" associations. There's more to it than that, but that's the basic idea.
	To a great extent, the false memories in the AMI have the properties of "real", "true" memories. For example, false recognition in the AMI is affected by depth of processing at the time of encoding, just like true recognition.
	And the AMI can induce priming effects affecting performance on stem-completion, fragment-completion, free-association, and other "implicit memory" tasks (McDermott, 1997).

	One interesting feature of the AMI is that, while it is strongly induced by pre-existing associative relations (hence the name), other strong associations don't necessarily induce the AMI. This was first noticed by Deese (1959) -- who, in addition to lists like the "needle" list, also employed a list of "associates" to the target fruit. But in this case, studying associates like apple, orange, and kiwi didn't induce false recognition of fruit.
	Park, Shobe, and Kihlstrom (1999) suspected that there was a difference in the effect of "horizontal" and "vertical" associations on the AMI. Most of the associates generated by needle are "horizontal", in that they remain at the same level of categorization. But most of the associates generated by fruit are "vertical", in that they are subordinate categories -- instances of fruit.
	Park et al. then performed an experiment systematically comparing lists of (horizontal) associates and (vertical) instances, controlling as much as possible for potentially confounding variables like associative strength. Subjects studied either lists of category exemplars, or lists of free associates to the category labels.
	The result was that the associative lists generated a strong AMI, while the categorical lists did not.
The conclusion from this experiment was that the AMI is a product of horizontal associative links, which stay at roughly the same level of categorization as the target. When the target is at a different level of abstraction than the stimuli, either the AMI doesn't occur at all, or subjects are able to recognize that sort of memory as false, and edit it out of their responses.

The Associative Memory Illusion is literally an illusion.

In the perceptual illusions, we perceive the present incorrectly, due to systematic distortions and biases. They are the product of constructive activity -- that the perceiver goes "beyond the information given" in the immediate stimulus environment.
In the illusions of memory, we remember the past inaccurately -- again, due to systematic distortions and biases. They are the product of reconstructive activity -- that the rememberer "goes beyond the information given" in the memory trace.

In the wake of the Bartlett Revival, we may say that the illusions of memory play the same theoretical role that the perceptual illusions played for Helmholtz, and for Rock, and Gregory, and Hochberg, and every other perceptionist who worked in the Helmholtz tradition. That is, they force us to the conclusion that the perceiver does more than unpack the stimulus. Rather, the perceiver makes his own contribution to the perceptual process, drawing on pre-existing knowledge, beliefs, and attitudes. So it is with memory, which brings us to...

A Unified View of Perception and Memory

An important corollary to the Reconstruction Principle is that memories are not properly viewed as traces of past events, encoded in the brain. Rather, they are better viewed as beliefs about the past.

The Reconstruction Principle also has important implications for the Library Metaphor of Memory, which has guided research and theory on memory ever since Ebbinghaus (if not Aristotle). It's convenient, and not completely incorrect, to think of memory as analogous to a book, which is written, cataloged and stored in a library, located, taken off the shelf, and read. But remembering is more like writing a book, based on fragmentary notes, than it is like reading one.

Remembering, like perceiving, is what Bartlett called effort after meaning. In both cases, the person is trying to make sense of his or her experience. Both perception and memory involve problem-solving activity -- the problem being to determine "What is going on now?" and "What was going on then?".

Perception is constructive activity, in which the perceiver builds a mental representation of the current environment -- a representation which, in some ways, may be illusory.
Memory is reconstructive activity, in which the rememberer builds a mental representation of the past -- a representation which, in some ways, may also be illusory.

Seven, Plus or Minus Two, Principles of Memory

The analysis of memory yields evidence of seven broad principles of memory, which may be organized according to the stage of memory processing to which they apply:

Encoding:

Elaboration
Organization

Storage:

Time-Dependency

I suppose I really ought to list Interference as the principle governing storage, because memory does not always fade with time (as Erdelyi's work on hypermnesia reminds us), and in any event the really important principle has to do with the mechanism underlying time dependency -- decay, displacement, consolidation, interference, etc. And since interference is the primary cause of forgetting from long-term memory (which is what I'm interested in), and because interference is consistent with the availability-accessibility distinction, I focus on interference rather than time-dependency. But I'm ambivalent about the matter, and I could well change my mind.

Retrieval:

Cue Dependency
Encoding Specificity

You might think that encoding specificity would supersede cue-dependency, but I keep them both in place because it was cue-dependency, and in particular the differing results of free recall, cued recall, and recognition tests, that underscored the distinction between availability and accessibility.

Schematic Processing
Reconstruction

There are other plausible candidates, and with all due respect to George Miller's famous paper, it's important not to make a fetish over the number 7.

As a reflection of the Bartlett Revival in the contemporary psychology of memory, Daniel Schacter has described "Seven Sins of Memory" that reflect the errors and biases entailed in reconstructive activity. These "sins" (Schacter himself admits that this might be too strong a word) can themselves be understood in terms of the "Seven Principles" discussed in these lectures.

In fact, there is one overarching principle, that runs through all of these, and underlies all of our understanding of how memory works (there, I think I've used all three possible spatial metaphors):

Availability vs. Accessibility.

Once encoded, memories are permanently available, and the problem is gaining access to them -- a problem that is solved, more or less successfully, by a variety of encoding and retrieval factors. (Maybe the availability-accessibility distinction should be added to the list of retrieval principles, but I think that cue-dependency and encoding specificity make the point).

It should be understood, however, that all of these principles were discovered in the context of standard tests of recall and recognition -- tests which require subjects to consciously remember past events. That begs the next question:

Is there more to memory than can be consciously remembered?

In one sense, that question has already been answered by the distinction between availability and accessibility. There is clearly more information available in memory than can be accessed at any particular moment, via any particular type of test.

But up until now, we have always defined access in terms of conscious recollection. Which begs the real question:

Can traces of past experience be encoded and stored, and thus available in memory, inaccessible to conscious recollection, but nonetheless still capable of influencing our ongoing experience, thought, and action?

The answer, it turns out, is "yes".

This page last modified 10/04/2014.