HookOne word made a car crash 27% faster
In 1974 Elizabeth Loftus and John Palmer showed 45 students the same film of a car crash, then asked one question with a single word swapped. 'About how fast were the cars going when they smashed into each other?' produced an average estimate of 40.8 mph. 'Hit' produced 34.0. 'Contacted' just 31.8 — a 27% swing off one verb. A week later, those given 'smashed' were more than twice as likely to 'remember' broken glass that had never appeared in the film.
That is the uncomfortable heart of this topic: memory is not a recording, it is a reconstruction, and it can be edited after the event. But before we can explain why it fails, we have to model how it works — the stores information passes through, the different kinds of long-term memory, and the mental workspace that holds material while we use it. This section runs from the architecture of memory to the reasons we forget and the reasons an eyewitness can send the wrong person to prison — and how a better interview can pull more truth out of the same witness. The exam reliably rewards one habit above all: quote the numbers.
ModelThe multi-store model — three stores, three specifications
Atkinson and Shiffrin (1968) pictured memory as a flow through three stores, each defined by coding (the form information is held in), capacity (how much) and duration (how long). The sensory register catches everything hitting the senses, coded by sense (iconic for vision, echoic for sound), with a huge capacity but a duration under half a second; only what we pay attention to passes on. Short-term memory (STM) codes mainly acoustically — Baddeley (1966) found acoustically similar words were confused in STM — holds about 7±2 items (Miller's 'magic number seven'; Jacobs found a digit span averaging 9.3 digits and 7.3 letters), and lasts roughly 18 seconds without rehearsal (Peterson and Peterson, 1959, using nonsense trigrams). Maintenance rehearsal keeps material in STM; prolonged rehearsal transfers it to long-term memory (LTM), which codes mainly semantically (by meaning), has a potentially unlimited capacity, and can last a lifetime — Bahrick et al. (1975) found 90% accurate recognition of school-mates' faces after around 15 years.
The model's weakness is its simplicity: patient KF (Shallice and Warrington, 1970) had a badly damaged STM yet an intact LTM, and only his verbal STM was affected, which a single unitary STM store cannot explain — and which the next two ideas dismantle.
ModelThree kinds of long-term memory
Tulving argued LTM is not one store but three. Episodic memory holds personal events, time-stamped and consciously 'travelled back' to — your first day at school, what you ate this morning. Semantic memory holds knowledge, facts and meanings that are not tied to a time or place — that London is the capital of England, what the word 'dog' means. Both are declarative: you consciously recall and can state them. Procedural memory holds skills and actions, recalled automatically and famously hard to put into words — riding a bike, driving a familiar route.
The evidence comes from clinical cases. HM and Clive Wearing could no longer form new episodic memories after brain damage, yet Clive Wearing still played the piano fluently (procedural intact) and both retained much semantic knowledge — proof the stores are separable and rely on different brain regions. For AO3, some psychologists (Cohen and Squire) argue episodic and semantic are really one 'declarative' system, and single case studies, however vivid, lack control and generalisability.
ModelThe working memory model — the mind's workspace
Baddeley and Hitch (1974) argued STM is not a passive store but an active system doing several things at once. The central executive is the attentional boss: it directs attention and allocates tasks to the subsystems, is 'modality-free', but has a very limited capacity. The phonological loop handles auditory and verbal information, coded acoustically; it splits into a phonological store (the 'inner ear' holding words heard) and an articulatory process (the 'inner voice' doing rehearsal), and holds only about two seconds of sound — hence the word-length effect. The visuo-spatial sketchpad is the 'inner eye' for visual and spatial data; Logie divided it into a visual cache (form and colour) and an inner scribe (arrangement and movement). The episodic buffer, added by Baddeley in 2000, is a general store that binds visual, verbal and time-ordered information into single episodes and links working memory to LTM.
Dual-task studies are the key support: two visual tasks interfere with each other, but a visual and a verbal task can be done together — exactly what separate subsystems predict. KF again fits (impaired verbal STM, intact visual STM). The recurring AO3 criticism is that the central executive is vague and hard to test directly.
Apply the model to a concrete scenario and it comes alive: a driver following a spoken sat-nav while watching the road is using the phonological loop for the directions and the visuo-spatial sketchpad for the traffic, with the central executive splitting attention between them — which is why they can manage both. But if a passenger starts giving verbal directions at the same time as the sat-nav, performance collapses, because both spoken streams now compete for the single phonological loop. An answer that walks through the scenario like this — naming each component and the task it is handling, then predicting when performance breaks down — is doing exactly the AO2 application the mark scheme wants, and it doubles as evidence for the model's central claim that STM has functionally separate stores.
MechanismWhy we forget — interference and retrieval failure
The spec wants two explanations for forgetting from LTM. Interference is forgetting because memories compete. In proactive interference, older learning disrupts newer — calling your new partner by an ex's name. In retroactive interference, newer learning disrupts older — your new phone number blocking recall of the old one. Interference is worst when the memories are similar: McGeoch and McDonald (1931) found recall was poorest when a second word list was made of synonyms of the first, and Baddeley and Hitch (1977) found rugby players recalled fewer teams the more games they had played, regardless of how much time had passed.
Retrieval failure (cue-dependent forgetting) says the memory is intact but inaccessible without the right cue — Tulving's encoding specificity principle: a cue helps only if it was present at both encoding and retrieval. Context-dependent failure: Godden and Baddeley (1975) had divers learn word lists on land or underwater, and recall was about 40% worse when the environment at recall did not match. State-dependent failure: Carter and Cassaday (1998) gave participants antihistamines, and recall was worse when the internal state at learning and recall did not match. For AO3, interference mainly explains a narrow range of similar material in artificial lab lists, and cues may not account for much everyday forgetting.
CaseEyewitness testimony — the memory the law relies on
Two factors distort what witnesses report. Misleading information comes in two forms. Leading questions: Loftus and Palmer's verb study (above), whose second experiment found 32% of the 'smashed' group but only 14% of the 'hit' group later reported seeing broken glass that was never there. Post-event discussion: Gabbert et al. (2003) found that when co-witnesses discussed an event, 71% went on to recall details they had not personally seen — memory conformity. Anxiety is the second factor. The weapon-focus effect: Johnson and Scott (1976) found witnesses who saw a man holding a bloody knife were less accurate at identifying him (about 33%) than those who saw him holding a pen (about 49%), because anxiety narrows attention onto the threat. Yet real-world evidence complicates this — Yuille and Cutshall (1986) found witnesses to a genuine Vancouver shooting stayed accurate five months later, and the most anxious were among the most accurate.
The resolution examiners like is the Yerkes-Dodson inverted-U: moderate arousal sharpens memory, while too little or too much degrades it — which reconciles the lab and field findings rather than treating them as a contradiction.
A model AO3 paragraph on misleading-information research shows the elaboration examiners pay for: 'A strength of Loftus and Palmer's research is its high degree of control: because every participant watched the identical film and only the verb changed, the difference in speed estimates can be confidently attributed to the leading question rather than to differences in what was witnessed, giving the study strong internal validity. However, this control comes at the cost of ecological validity — watching a filmed crash in a lab carries none of the shock and consequence of witnessing a real one, so participants may be more suggestible than genuine eyewitnesses. This matters because Yuille and Cutshall found that witnesses to an actual shooting resisted misleading questions, suggesting lab studies may overstate how unreliable everyday testimony really is.' The paragraph never just labels a flaw. It states the point, explains the mechanism behind it, then delivers a consequence that reaches beyond the study to the real-world question the law actually cares about. That final 'so what' — weighing lab control against field validity — is the analytical step that lifts an answer into the top band.
MechanismThe cognitive interview — pulling more truth from the same witness
If ordinary questioning distorts memory, a technique built on memory theory can improve it. Fisher and Geiselman's cognitive interview uses four instructions. Report everything — include every detail, however trivial or partial, because small fragments may cue larger ones. Reinstate the context — mentally return to the scene, its weather, sounds and your feelings, exploiting context-dependent retrieval. Reverse the order — recall events in a different sequence to disrupt schema-driven expectations of how things 'should' have gone and to make fabrication harder. Change perspective — describe the scene from another person's viewpoint to break up the effect of schemas. The enhanced cognitive interview adds social elements: building rapport, minimising distractions, and letting the witness control the pace.
Köhnken et al.'s (1999) meta-analysis of 55 studies found the cognitive interview produced about 34% more correct information than a standard interview — though also more incorrect details, a trade-off police must manage. For AO3, it is time-consuming and needs trained officers, and the elements are not equally useful: Milne and Bull found that 'report everything' combined with 'reinstate the context' was the most effective pairing.
VocabularyKey terms the mark scheme pays for
TrapsMisconceptions that cost marks
ExamWhat examiners want
AQA marks AO1 (describe), AO2 (apply to the stem) and AO3 (evaluate); the 16-mark essays are 6 AO1 and 10 AO3, so evaluation carries the grade. For AO1, quote the specification with figures — '7±2 items', '18 seconds', 'about 40% worse' — because the mark scheme separates accurate, detailed knowledge from vague description. Learn the coding/capacity/duration table cold, because short-answer questions test it directly.
Case studies are your evaluation currency: KF challenges the multi-store model's single STM, HM and Clive Wearing support the split between episodic, semantic and procedural memory, and you should state what each case shows and the standard counterpoint (case studies lack control and generalisability). On eyewitness-testimony essays, the top-band move is to weigh lab studies (high control but low ecological validity — Loftus and Palmer) against field studies (real validity — Yuille and Cutshall) rather than listing them. On application questions, name the exact mechanism (proactive interference, context-dependent retrieval, the phonological loop) and tie it to the specific detail in the scenario — an answer that never refers back to the stem is capped regardless of how much theory it contains.