The hippocampus and the amygdala: what each does for memory

Where are they?#

Both sit in the medial temporal lobe, the inner part of the temporal lobe. Simply Psychology’s rules of thumb are that the hippocampus is just above each ear and that the amygdala is roughly level with your eyes.

Both are usually counted in the limbic system. This page is about how they normally work. Apart from H.M.’s place in the history, it doesn’t cover memory problems, and it gives no health advice.

What does the hippocampus do?#

It’s needed to form new memories of events and facts

These are declarative memories, the kind you can state, which take in both episodic and semantic memory. Squire and colleagues put it this way: conscious memory for a new experience is “initially dependent on information stored in both the hippocampus and neocortex”.

What the hippocampus holds may be closer to a set of pointers. Hippocampal memory indexing theory, from Teyler and DiScenna in the 1980s, says it “does not ‘contain’ the episodic memory itself”. It makes a code that binds the activity patterns of the experience, which are stored across the cortex. Goode and colleagues open their review of the theory with Google’s index as the comparison.

It guides a slow reorganization

In Squire and colleagues’ standard account, systems consolidation is the process by which the hippocampus “guides the reorganization of the information stored in the neocortex” until that information no longer needs it, over weeks, months or longer. Whether it ever fully stops needing the hippocampus is disputed, as covered below.

One proposed means is replay, where hippocampal activity from learning recurs by itself. It’s been seen most often in rats during slow-wave sleep, and replay-like activity has been seen during quiet waking too. The Field Guide’s page on consolidation goes further into this.

It helps you find your way

In 1971 John O’Keefe found cells in the rat hippocampus that fired when the animal was in one place in a room, and other cells that fired in other places. In 2005 May-Britt and Edvard Moser found grid cells next door in the entorhinal cortex. The three shared the 2014 Nobel Prize, and the committee notes evidence that both cell types exist in people.

The best-known human study is of London taxi drivers. Maguire and colleagues found the rear of the hippocampus larger, and a more forward part smaller, in licensed drivers than in people who didn’t drive taxis, and the difference tracked time in the job. That can’t show which came first, so Woollett and Maguire later scanned trainees before and after their 3 to 4 years of study. Those who qualified gained gray matter in the rear of the hippocampus.

Whether memory or space is the core job hasn’t been settled. Simply Psychology sets out the two camps and Cohen and Eichenbaum’s relational view, in which the hippocampus handles flexible relations of many kinds, such as the order of events. Anand and Dhikav say it’s also been implicated in imagining the future.

Inside it: a loop

The parts have names students get asked about. The usual diagram is a loop that starts and ends at the entorhinal cortex.

Pattern separation keeps similar experiences apart, this morning’s breakfast from yesterday’s. Pattern completion rebuilds a whole from a fragment. CA stands for cornu ammonis.

What does the amygdala do?#

It learns what predicts a threat

In fear conditioning, a neutral tone is paired with a shock until the tone alone sets off a defensive response. Much of this work is in rats. Squire lists emotional learning among the kinds of memory outside the hippocampal system, with the amygdala important for it.

It strengthens what’s stored elsewhere

This is a separate job, and the one most tied to everyday memory. James McGaugh’s account runs like this: emotional arousal releases the adrenal stress hormones epinephrine and cortisol, those act through the amygdala, and the amygdala’s projections to other regions strengthen the consolidation going on there. Much of it comes from rats, where drugs delivered to the amygdala after training raised or lowered later memory.

The human evidence includes two studies with very different methods:

• Cahill and colleagues scanned people watching emotional and neutral film clips. Three weeks later, activity in the right amygdala during the emotional films was highly correlated with how many of them people recalled. It wasn’t significantly correlated for the neutral ones.
• Inman and colleagues worked with 14 epilepsy patients who already had electrodes in place for their care. One second of mild stimulation of the amygdala followed half of a set of neutral pictures, and people reported no emotional response. Memory for those pictures wasn’t reliably better right away, and it was the next day.

Higher means better at telling old pictures from new. The gap on the immediate test wasn’t reliable (P = 0.30), and the next-day gap was (P = 0.003). That delay is what you’d expect if the effect is on consolidation, and the authors say the amygdala may be able to prioritize memories without an emotion being felt.

Mild emotion is enough. McGaugh reports better memory for material only mildly emotional, positive or negative, and for word lists followed by an arousing video, one comedy and one of surgery.

Vivid isn’t the same as accurate

Sharot and colleagues scanned 24 people who were in Manhattan on September 11, 2001, about three years later. Those who had been downtown, close to the attacks, showed selective amygdala activation when recalling that day and rated those memories as more vivid and relived. Those further away in Midtown didn’t.

The study didn’t test accuracy. Its review of earlier work says these memories “are not especially accurate” and are held with great vividness and confidence.

How do they work together?#

In the near miss, the hippocampus links what you saw and heard into one event. Arousal engages the amygdala, which turns up how strongly that gets kept. In the 5 of Inman’s patients with electrodes in all three regions, the next-day test showed stronger interactions among the amygdala, hippocampus and perirhinal cortex for stimulated pictures they remembered, which the authors read as the amygdala engaging other memory regions.

Threat learning is where the two come apart most cleanly. LeDoux summarizes the patient studies: hippocampal damage “disrupts explicit conscious memory of having been conditioned but has no effect on fear conditioning itself”, and amygdala damage “disrupts fear conditioning but not the conscious memory of having been conditioned”. So one person reacts to the tone with no memory of the training, and another remembers the training and doesn’t react.

Who was H.M., and why does every textbook mention him?#

Henry Molaison, known in papers as H.M., had surgery in 1953 at 27 to control seizures that had left him unable to work. William Scoville removed tissue from the medial temporal lobe on both sides. The seizures came under control, and Brenda Milner found a man who, in Larry Squire’s words, “forgot daily events nearly as fast as they occurred”, with no general intellectual loss. Scoville and Milner reported him in 1957, and he was studied for five decades until he died in 2008 at 82.

His case set the field’s starting points. Memory could be separated from intelligence and perception. He could hold a conversation and repeat back six or seven digits, and lost the material once his attention moved on, which supported the distinction between short-term and long-term memory. Being unable to form new long-term memories is what’s called anterograde amnesia. Early tests found his older memories largely kept. How far back the loss ran, and whether detailed episodes survived, has been argued over. He learned to trace a star in a mirror and kept the skill across three days, with no recollection of having done it, which was the first hint of more than one kind of memory. Procedural memory has that task as a demo.

By the surgeon’s description, the operation included the hippocampus, the amygdala and the neighbouring parahippocampal gyrus, so it couldn’t say which part mattered. The 1957 paper ends by saying it isn’t known whether the amygdala plays any part. The damage also turned out to be less extensive than first described. Later work pinned the memory system to the hippocampus and the cortex around it. Squire writes that damage limited to the hippocampus “causes significant memory impairment but considerably less impairment than in H.M.”

What do people get wrong about them?#

That H.M. simply had his hippocampus removed

The Queensland Brain Institute says he “had his hippocampus surgically removed”. Squire reports that the early paper is “sometimes cited incorrectly” as showing the hippocampus matters for memory, since the operation included more, and he quotes Milner: “I have never claimed that the memory loss was solely attributable to the hippocampal lesions”. Simply Psychology’s hippocampus page also lists the amygdala and entorhinal cortex among what was removed.

That the hippocampus is the memory center, or that it’s settled what old memories need

Simply Psychology’s amygdala page calls the hippocampus “the brain’s memory center”, while its hippocampus page calls that label “the most persistent misconception”, and says old memories “survive its loss because they no longer need it”.

That second line is one side of an open argument, as the same page says further down. Multiple-trace theory holds that memories with contextual detail stay dependent on the hippocampus for as long as they last. Squire’s review, from the other side, calls the patient findings on old autobiographical memories mixed, says damage outside the medial temporal lobe may explain part of that, and concludes that such memories “ultimately depend on a distributed network of cortical regions”. We’ve read that account of the dispute and not the theory’s authors.

That the amygdala is the fear center

Simply Psychology notes that the amygdala is “often introduced as the brain’s ‘fear centre,’” and says the label oversimplifies. A researcher whose work helped build that reputation has backed away from the word. Joseph LeDoux wrote in 2014 that calling the threat-detecting system the fear system “was a mistake that has led to much confusion”, because those circuits are distinct from the ones that give rise to the conscious feeling.

In 2013 Feinstein and colleagues reported that three people with damage to both amygdalae felt fear and had panic attacks after breathing air with 35% carbon dioxide, a threat from inside the body. The paper’s main text words the finding narrowly: “the amygdala is not required for fear and panic evoked by CO₂ inhalation”. The study separates that from fear of outside threats, and where measured, the patients’ responses in anticipation of the inhalation were deficient. Adolphs’s review: “There is no single brain structure for processing fear”.

That an “amygdala hijack” is an established scientific term

It’s a popular phrase, coined by Daniel Goleman in his 1995 book Emotional Intelligence. Wikipedia says it’s “recognized as a formal academic term within affective neuroscience”, and we found no support for that. A search of titles and abstracts on Europe PMC on 1 November 2026 returned two papers: one on conflict resolution that names it as a risk, and one on fear during COVID-19, which has no abstract and which we didn’t read. Neither Pessoa and Adolphs nor McFadyen and colleagues, below, use the phrase, and neither tests a hijack.

Simply Psychology and Wikipedia explain it with a fast route from the thalamus that reaches the amygdala before the cortex can weigh in. In rodents, a route like that carries sounds of threat. For vision in humans, Pessoa and Adolphs review the anatomy and physiology and “argue against the notion” that a subcortical route plays a prominent part, and give the cortex a larger role. McFadyen and colleagues reconstructed such a pathway by tractography in 622 people in 2019, linked its fibre density to fear recognition, and still call the continuity of the route disputed.

Do AI systems have anything like a hippocampus or an amygdala?#

Not as structures, though both have been borrowed from as ideas. The closest match to the hippocampus is the offline pass some AI memory systems run, merging recent notes into longer-lasting ones, which the Field Guide traces back to consolidation. A search index over stored text is a loose match for indexing theory. It points at the pieces, and nothing in it does what the hippocampus does while an event is happening.

For the amygdala, most systems keep a single importance score. Counterparts, the memory system whose site this is, records feelings when a memory is written and lets the strongest one slow that memory’s fading. That’s a design choice inspired by this work, and the human studies above measure how much is remembered later, which doesn’t settle a rate of fading. The directory compares the systems.

Questions people ask#

Questions about how any of this carries over to AI? Ask in r/AIMemorySystems.