Do I still need to remember things?

For most facts, no, and fifteen years of offloading research gives you permission rather than guilt. For a narrow set, yes, and the reason is not nostalgia. Memory is what lets you notice that an answer is wrong. A model producing something confident, plausible and false is caught only by a person carrying enough of the subject internally to feel the friction. Everything else is retrievable. That reflex is not.

Offloading has a documented upside, and it gets left out#

The alarming version of this literature is the only version most people meet, so start with the finding that runs the other way.

Storm and Stone reported three experiments in Psychological Science on what they named saving-enhanced memory. Saving one file before studying a new file significantly improved memory for the contents of the new file. Offloading the first thing freed capacity for the second.

Two conditions broke it, and both matter. The benefit disappeared when the saving process was deemed unreliable, and when the contents of the saved file were not substantial enough to interfere with the new material in the first place.

The first of those has a consequence people rarely draw. If you do not trust the store, offloading stops working as a memory strategy, because part of your attention stays behind guarding the thing you supposedly put down. The reliability of the external system is therefore not a separate question from your own cognition. It is inside it. Which is an uncomfortable thing to notice about a store that sometimes invents its contents.

This page states the Storm and Stone result without figures. The full text sits behind a paywall and could not be read at the primary source, so only what the published abstract asserts is reported here. That is a weaker basis than the rest of this page and is flagged rather than smoothed over.

What expected availability actually changes#

Sparrow, Liu and Wegner ran four experiments in 2011 and found that when people expect information to remain available, they remember where to find it rather than the thing itself. The finding is about a shift in what gets encoded.

The authors are careful about what it does not show, and the caution has been widely discarded on the way to the headline: it does not establish that total memory capability declines, or that the trade is net negative. Human beings have always distributed memory into other people, notebooks and institutions. A machine that answers is a new store, not a new behaviour.

Risko and Gilbert's review supplies the part that should worry you more. People offload not only when a task is genuinely hard but when they judge it to be hard. The decision is metacognitive and frequently mistaken, which means the boundary of what you have stopped remembering is not being set by any deliberate policy of yours.

The one place a capability cost has actually been measured#

Dahmani and Bohbot studied habitual satnav use with a cross-sectional comparison and a three-year longitudinal follow-up. Heavy users had worse spatial memory when navigating unaided, and heavier use over the following three years was associated with steeper decline.

This is the strongest everyday evidence that a reliably performed external function is associated with weakening of the human equivalent over years. It is also correlational, and it concerns spatial memory rather than reasoning, both of which the study's own framing acknowledges. People who dislike navigating may simply use satnav more.

Hold it next to the clinical finding on the same estate: experienced endoscopists whose unassisted detection rate fell from 28.4 to 22.4 per cent within months of AI exposure. Different domain, same shape. The loss is invisible while the tool is present and appears only in its absence, which is the pattern this research calls capability debt.

Why the loss is hard to feel#

Fisher, Goddu and Keil ran nine experiments showing that searching the internet for explanations inflates people's ratings of their own ability to explain unrelated topics, with effect sizes between d = 0.35 and d = 0.63. The illusion persisted when the search returned no answer to the question asked, and when it returned no results at all. It vanished for autobiographical topics, where searching would not help, so it is targeted rather than general.

Their own summary of the mechanism is the useful sentence: it is not that people misattribute where the knowledge came from, but that they inflate the sense of how much of the total is stored internally.

Which means the self-check most people would apply here, asking yourself whether you still know things, is measuring the wrong quantity with a broken instrument. The only reliable test is to try to do it without the tool, which is the diagnostic set out at am I becoming dependent on AI.

Three things worth keeping in your own head#

Not a nostalgic list. Each of these earns its place by being unavailable at the moment you need it.

• The structure of any domain you are accountable in. Not the facts, the shape: what depends on what, what the usual ranges are, what would be surprising. You cannot recognise a wrong answer in a field whose shape you do not hold, and recognition is what you are being paid for once production is cheap.
• The failure patterns of your own field. Error detection is largely pattern recognition for things that have gone wrong before. This is the part that is genuinely hard to reacquire, because it was built by encountering the errors rather than by reading about them.
• Anything where the delay would change the decision. Retrieval is free and it is not instant. In a conversation, a negotiation, a consultation or an argument, the fact you have to look up has already arrived late.

Everything else can go into the store with a clear conscience. Phone numbers, dates, procedures you follow rather than judge. The historical anxiety about writing destroying memory was wrong about writing, and most of the current anxiety is wrong in the same way.

What nobody has shown#

• That generative AI produces the satnav effect for reasoning. Dahmani and Bohbot is spatial memory and correlational. The extension to knowledge work is an inference and is labelled as one everywhere on this site.
• That the Sparrow effect is harmful. Its authors say it is not established, and fifteen years later it still is not.
• Whether saving-enhanced memory holds for an unreliable store. Storm and Stone found the benefit disappeared when saving was deemed unreliable, but nobody has run the equivalent test with a system that is usually right and occasionally fabricates, which is the actual condition people are in.
• Whether any of this reverses. The recovery question is treated separately at can you regain a skill you have lost, and the answer there is thinner than anyone would like.
• What happens over a lifetime. Every study here runs from minutes to three years. Nobody has followed a person from twenty to sixty.

Key research and primary sources

• Storm, B. C. and Stone, S. M. (2015). Saving-Enhanced Memory: The Benefits of Saving on the Learning and Remembering of New Information. Psychological Science, 26(2).
• Sparrow, B., Liu, J. and Wegner, D. M. (2011). Google Effects on Memory. Science, 333(6043).
• Risko, E. F. and Gilbert, S. J. (2016). Cognitive Offloading. Trends in Cognitive Sciences, 20(9).
• Dahmani, L. and Bohbot, V. D. (2020). Habitual use of GPS negatively impacts spatial memory during self-guided navigation. Scientific Reports, 10, 6310.
• Fisher, M., Goddu, M. K. and Keil, F. C. (2015). Searching for explanations. Journal of Experimental Psychology: General, 144(3).
• Budzyn, K., Roman'czyk, M., Kitala, D. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology and Hepatology, 10(10).

Related SuperSkills research#

On the concepts, cognitive offloading, the Google effect and deskilling. On the diagnostic, am I becoming dependent on AI and using AI without dependency. On recovery, can you regain a skill you have lost. On the reading version of the same question, should I let AI summarise everything I read. On what a machine remembering you changes, should AI remember everything about me.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Sparrow, Risko and Gilbert, Dahmani and Bohbot, Fisher and Budzyn were read at the primary source and their figures checked. Storm and Stone is paywalled and only its published abstract could be read, so this page reports that paper's qualitative claims and gives no figures from it.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.