Serial Recall Effect: Why Order Slips in the Middle of a Sequence

Remembering a sequence means remembering two things, the items and their order, and the order is the part that fails first. People asked to repeat nine words will often produce eight or nine of the right words with two of them swapped. The errors are not random: they follow the same shape across words, digits, letters and spatial positions, and that shape is what the serial recall effect names.

The first feature is a bowed curve by position. The opening items are recalled in place most often, the last one or two gain a little, and the middle is worst. It resembles the curve for free recall of a list, but the balance is different: in order recall the advantage of the start is large and the advantage of the end is small, because each item recalled out of place in the middle disturbs the ones that follow it.

The second feature is locality. When an item is reported in the wrong position it is usually one place away from where it belonged, and the probability of an error falls steeply with distance. Henson, Norris, Page and Baddeley documented this in 1996 and set it alongside two further regularities: items that sound alike tend to swap with each other, and repeating an item you have already produced is rare.

These regularities decide between theories. Chaining models, in which each item is a cue for the next, predict that a wrongly recalled item should derail the item after it. Henson and colleagues built sequences that alternated similar-sounding and distinct items and found the similar items were recalled worse, producing a sawtooth curve, while the distinct items between them were untouched, which chaining cannot explain. Positional models can. In Henson's 1998 Start-End Model, every item is tagged with its distance from the start and from the end of the sequence. Items near either anchor have sharp tags and are rarely displaced; items in the middle have similar, blurry tags and swap with their neighbors. Hurlstone, Hitch and Baddeley's 2014 review found the same principles, a primacy gradient of activation, position marking, and suppression of items already produced, across verbal and spatial sequences.

Similarity of sound matters because the tags for order are not the only thing stored. Conrad and Hull showed in 1964 that sequences of letters that sound alike are harder to repeat, and Baddeley's 1968 experiments located the damage in retrieval, as confusions between items rather than loss of them.

The pattern was mapped over decades rather than discovered once, and the two papers this page leans on are Richard Henson's. In 1996, with Dennis Norris, Michael Page and Alan Baddeley, he published "Unchained memory: Error patterns rule out chaining models of immediate serial recall" in the Quarterly Journal of Experimental Psychology. Participants saw or heard short sequences and repeated them in order. Sequences that alternated confusable and non-confusable items produced more errors on the confusable items, a sawtooth across positions, while the presence of a confusable item did not increase errors on the distinct item that followed it and did not change the kind of error made. The authors then analyzed every error: most transpositions were over short distances, confusable items tended to interchange, and repeated responses were rare and far apart. They named these the locality, similarity and repetition constraints and argued that they ruled out chaining as the mechanism of serial order.

In 1998 Henson published "Short-term memory for serial order: The Start-End Model" in Cognitive Psychology. The model codes each item's position relative to the start and end of the sequence and was shown in simulations to reproduce the main phenomena of serial recall: the effects of primacy, recency, list length, grouping, modality, a redundant suffix, proactive interference, retention interval and phonological similarity. It was also the first model to capture the full pattern of errors, including transpositions, repetitions, omissions, intrusions, confusions and the positional errors that cross group boundaries and even trials, and it predicted correctly that an item displaced into another group keeps its relative position within that group.

Earlier work supplied the pieces. Conrad and Hull's 1964 paper on acoustic confusion and memory span showed that similar-sounding items shrink the span, and Lee and Estes developed in 1981 the account on which order information drifts, or perturbs, over time so that neighboring items exchange places, which is the ancestor of the locality constraint. The demonstration on this page uses three position probes rather than full ordered recall, so it can show the shape of the pattern without being able to measure it in you.

The rows below are illustrations of where order has to be held briefly in memory, not findings from the studies cited on this page.

Order memory cannot be trained out of these errors, but the sequences you are given can be shaped so the errors matter less.

• Group the sequence. Grouping into threes or fours gives every group its own start and end, so more items sit next to an anchor and fewer sit in a blurry middle.
• Write it down or read it back. A sequence held for ten seconds in memory is at the mercy of the curve; a sequence on paper is not.
• Avoid similar-sounding items in codes you design, and spell out confusable letters with a phonetic alphabet when you read codes aloud.
• Put the item that must not move at the start or the end. The middle positions are where swaps happen.
• When order matters for a procedure, number the steps and tick them off, rather than carrying the order in your head between steps.
• When taking an account of events, record the order as it is given and treat a one-step disagreement about the middle as the expected error rather than as a sign of dishonesty.

The serial recall effect is about memory for order. It is easy to confuse with findings about which items are remembered, and with findings about how many.

• It is not the serial-position effect, the U-shaped curve of free recall in which the first and last words of a list come back and the middle is lost. That effect concerns whether an item is recalled at all, in any order; this page concerns whether an item is recalled in the right place, and in order recall the ends are protected for a different reason, their sharp positional tags.
• It is not the magical number seven, the limit on how many items can be held at once. Span sets the length at which order errors begin; the pattern described here is how those errors are distributed once the limit is reached.
• It is not the suffix effect, where a redundant spoken item after the list removes the advantage of the final position. That is one of the phenomena the Start-End Model reproduces, but it concerns the last item only.