Remembering Where to Look

The moment you type something into a notes app, your brain starts forgetting it. Not immediately — but the moment you register that it's captured somewhere, the internal encoding process quietly winds down. This isn't laziness. It has a name.

In 2011, cognitive psychologist Betsy Sparrow at Columbia ran a series of experiments with a simple setup: tell people they can look something up later, and watch what happens to their memory. The results were consistent and a little uncomfortable. People who expected to find information in a file later remembered the information itself less well — but they remembered where to find it more accurately than people who had no such expectation. The paper landed in Science. Sparrow and her co-authors Jenny Liu and Daniel Wegner called it the Google Effect.

The uncomfortable interpretation is that Google made us dumber. The more honest one is that this has been happening for thousands of years.

Offloading is the strategy, not the failure

Wegner had described the basic mechanism back in 1985, before anyone had a smartphone. He called it transactive memory: the system by which people in close relationships divide up the cognitive labor of knowing things. One partner tracks which friends are having trouble. The other knows which turn to take at the highway interchange. Neither tries to hold both. That's the point.

The library card catalog was transactive memory. So was the address book in your grandmother's kitchen drawer. The human brain has always treated other people and external systems as legitimate memory partners. What Sparrow's research added is that we're quick to extend this arrangement to new partners — even ones we just met, including software.

What makes a good transactive partner isn't how much it stores. It's how reliably you can get the information back when you need it.

The part that breaks

When Wegner studied divorced adults, one finding was quietly striking: people in the aftermath of a relationship breakdown often described a period of cognitive disorganization that went beyond the emotional. Functional confusion. They had spent years routing half their mental model of the world through their partner. That routing was gone. The information hadn't disappeared — but the pointer had.

The same thing happens when a notes app discontinues. Or when Evernote migrates your data in a way that breaks your folder structure. Or when you switch email clients and discover you'd been relying on search that worked better than you knew.

A bad transactive partner is worse than no transactive partner, because you encode less in your own head and then can't retrieve it from the external store either. You get the forgetting without the recall.

This is the failure mode that AI memory systems are quietly building toward.

What AI memory gets wrong

ChatGPT's memory system, which rolled out to free users in mid-2025, works roughly like this: the model accumulates facts about you across conversations, stores them somewhere, and applies them to future responses. You can view the list in theory. In practice, most users don't, and it's not obvious how to reason about what's in there. Charles Packer, who has spent years building AI memory infrastructure, put it plainly: "the biggest problem with OpenAI's ChatGPT memory system is that it's a black box."

So the situation is: you encode less carefully because you trust the system is capturing things. But you can't inspect what it has. You can't tell whether the fact it stored eight months ago is still true. You can't search through it the way you'd search through a folder of notes. The transactive partner is accumulating state about you, and you have limited visibility into what that state actually is.

You've outsourced the encoding. The retrieval is opaque. This is worse than just forgetting, because at least when you forget, you know you don't know.

What visibility changes

The reason structured, inspectable memory matters isn't about control as a philosophical value. It's because the transactive memory effect is real and immediate: we do stop encoding as carefully once we trust our external stores. If the external store is unreliable or uninspectable, you pay double. You didn't encode it carefully. Now you can't find it either.

The opposite design earns a different kind of trust. Not just psychological trust but epistemic trust — the kind where you can look at what the system knows and ask whether it's right. Where you can find the note from last September. Where you can correct a stale preference before it propagates into another six months of slightly-off responses.

People have been building transactive memory systems long before computers. Commonplace books, well-indexed filing cabinets, address books organized by something other than alphabetical first name. The best ones shared a property: you knew what was in them. Not because you held it all in your head, but because the organizational logic was visible enough to navigate.

---

An AI that silently accumulates facts about you is a transactive partner that hasn't told you what it remembers. The cognitive cost isn't theoretical — it's baked into how human memory has worked as long as we've been building external stores for it. We offload. We always have. The question is what we're offloading to, and whether it can be trusted to hold up its end of the deal when we actually need it.

We've always been remembering where to look. We just haven't always chosen the right places to look.

---

Asgeir Albretsen is the founder of Harbor.