What You Trade When AI Does the Remembering

In 2000, Eleanor Maguire at University College London scanned the brains of London taxi drivers and found something unexpected: their posterior hippocampi were significantly larger than those of matched controls. Not slightly, but measurably, structurally different. The difference correlated with years on the job. The longer a cabbie had spent learning "the Knowledge" (London's licensing requirement, which means memorizing 25,000 streets and everything along them) the more their brain had physically adapted to hold it. By 2011, follow-up work showed the hippocampal advantage was shrinking. GPS had arrived. Fewer people learned the Knowledge the hard way. The brain, it turns out, doesn't maintain structures it no longer needs.

This is the most vivid example I know of cognitive offloading carrying a measurable cost. But there's something easy to miss in the obvious reading: the taxi drivers weren't losing a skill to laziness. They were losing it to a better tool. GPS is genuinely better at navigation than memory. No cabbie navigates more accurately by refusing to use it. The loss is real. The loss is probably worth it. The question is whether you knew what you were trading.

The case for offloading

Cognitive offloading is as old as writing. The first time someone carved information into a clay tablet rather than memorizing it, they were offloading. Plato, famously, thought this was catastrophic. Writing, he argued through Socrates in the Phaedrus, would weaken memory and create the mere appearance of wisdom without the substance. He was partly right and almost entirely wrong.

Writing made civilization possible. Calendars made coordination possible. Filing cabinets, spreadsheets, contact lists: each offloads something from internal memory to external storage, and almost every trade has been worth making. Human cognition didn't shrink because of the printing press; it redirected.

What the GPS research actually shows isn't that tools are bad. It's that specific cognitive skills atrophy when you stop practicing them. The hippocampal difference in London cabbies wasn't general intelligence. It was a specific spatial navigation skill, built through intensive practice, maintained precisely because it was being used. Stop using it, lose it. That's fine if you don't need it. The question that matters is whether the skill being offloaded is one you actually wanted to keep.

The part that's less obvious

A 2025 MIT Media Lab study gave 54 participants essays to write with varying levels of AI assistance, over four months. EEG data showed that participants who relied most heavily on AI exhibited systematically weaker neural connectivity than those who worked alone or used only search. More striking: 83% of the AI-assisted group couldn't accurately recall what was in the essays they'd just written with help. The study is preliminary, not peer-reviewed, and has a small sample. But the pattern it describes is plausible.

The effort of thinking through a problem, including the friction of retrieval, is part of how memory encodes and skill develops. Psychologist Robert Bjork calls these "desirable difficulties": cognitive obstacles that feel inefficient but produce more durable learning than frictionless alternatives. The same logic explains why testing yourself outperforms re-reading notes, and why writing by hand tends to produce better recall than typing. The generation of effort isn't a bug in the learning process. In many cases it is the process.

Remove the struggle, and you might be removing the thing that made the knowledge stick.

Three kinds of offloading

Not all offloading carries the same risk. Offloading storage is usually fine. Writing something down instead of memorizing it frees working memory without degrading much. Humans have been doing this for five thousand years without obvious intellectual decline. The clay tablet people seemed to manage.

Offloading retrieval is more complicated. When you struggle to remember something and successfully retrieve it, that struggle strengthens the memory trace. When a system retrieves it for you instantly, the strengthening doesn't happen. You get the answer, but you don't build the recall path. Over time, across enough questions, that compounds into something.

Offloading processing is where most of the risk lives. When AI summarizes, synthesizes, and concludes, it's not assisting your thinking. It's replacing the thinking. That's qualitatively different from writing a note or querying a filing system.

Where this leaves me

I built Harbor partly as a storage offloader. The premise is that context you've written down is more useful than context you're trying to hold in your head, especially when AI tools can retrieve and query it. That premise still seems right to me.

But I'm less certain that frictionless retrieval is cost-free. If the act of searching your own notes, actually looking through them rather than asking an AI what they say, is also the act of consolidating what you know, then optimizing that away might feel cheaper in the short run and cost more over time.

I don't think the answer is to make tools deliberately worse. And I don't think most people are in serious danger of losing important cognitive skills because they use AI assistants for notes. The London taxi driver scenario required years of intensive spatial practice followed by years without it. Most knowledge work doesn't build that kind of dedicated neural real estate to begin with.

But the GPS research found that people who used turn-by-turn navigation regularly had worse spatial memory even after controlling for prior navigation ability. The effect isn't limited to experts who've trained exceptional faculties. Habitual offloading changes the average person too.

The honest answer is that I don't know exactly where the line is. I'm fairly confident it's not at "write things down." I'm less confident it's not at "never struggle to recall anything because your tools will answer instantly." Somewhere between those two points, the tradeoff shifts from worth it to something else. Finding that line seems like it's worth paying attention to, especially right now, when the tools are getting fast enough that the friction disappears before you notice it was doing anything.

---

Asgeir Albretsen is the founder of Harbor.