Why the Best Memory Forgets

In 1942, Jorge Luis Borges wrote a short story about a man who couldn't forget anything. Ireneo Funes, after being thrown from a horse, woke with perfect memory. Every leaf he'd ever seen. Every cloud. Every moment. He remembered each thing with such precision that he found the very concept of categories offensive — how could you call two dogs by the same word when they were so manifestly different? He lay in a dark room, overwhelmed. He died at nineteen.

"To think is to forget differences," Borges writes. "It is to generalize, to abstract."

Funes couldn't think. He could only remember.

The neuroscience version

In 2017, Paul Frankland and Blake Richards at the University of Toronto published a paper in Neuron that, had Borges been alive, he would have found satisfying. Their argument, backed by decades of memory research, was that the goal of memory is not to transmit accurate information over time. The goal is to guide intelligent decision-making. And for that purpose, forgetting isn't a failure of the system — it's one of the system's most important features.

The mechanism is specific. As new neurons form in the hippocampus and integrate into existing circuits, they overwrite older memories. This isn't an architectural flaw; it's how the brain makes room for generalization. When you remember the principle rather than the instance, you can apply it to new situations. The brain that forgets exactly which Tuesday you read something is the same brain that retains the insight from it.

Frankland and Richards called this "transience" — the complement to "persistence" — and argued that both are required for intelligent behavior in changing environments. Perfect memory, paradoxically, makes you worse at prediction.

The note-taking version

The PKM community discovered a softer version of the same truth. Tiago Forte built an entire methodology — progressive summarization — around the idea that information becomes more valuable as it gets compressed. You highlight a document. Then you highlight your highlights. Then you write a summary. The layers of reduction are the point. What you end up with isn't a perfect record of the source; it's the part that mattered to you, in your words, at the level of abstraction you needed.

Nobody says this plainly, but progressive summarization is a theory of intelligent forgetting. You're not trying to preserve everything. You're deciding what survives.

The failure mode of most note-taking systems is the opposite: they optimize for capture and neglect selection. Notion databases with thousands of entries that nobody searches. Readwise highlights that pile up faster than anyone reads them. An inbox that grows faster than it empties. More storage, more noise, less usefulness.

A note you can't find is worse than no note at all. At least with no note, you go look the thing up.

What this means for AI memory

The pitch for AI memory is almost universally framed as: the AI will remember more things about you, so it will be more helpful. Everything you said. Everything you wrote. All of it, indexed and retrievable. More context, better responses. The logic seems obvious.

But Frankland and Richards point somewhere else. Memory that serves intelligence isn't comprehensive; it's selective. What matters isn't whether you can retrieve any given fact, but whether the facts you retrieve help you make good decisions in the current situation.

An AI that retains everything has the Funes problem. It doesn't know which memories are relevant. It surfaces details when you need principles, and principles when you need details. The past drowns the present.

The alternative isn't less memory — it's structured memory. Typed entities. Explicit relationships. Things stored not as a blob of undifferentiated text but as records with kinds and dates and sources. When an AI can tell the difference between a fact about someone's job title (probably stable) and a note about their current project status (probably stale), it can reason about which to surface and when. That's not a minor improvement in recall. It's a different theory of what memory is for.

The hard part

Selecting what to keep is harder than recording everything. Recording is passive; selection requires judgment. And judgment is uncomfortable to delegate.

It's more satisfying to believe that if you capture enough, the useful thing will surface when you need it. The idea that you have to decide — at time of capture — what's actually worth keeping feels like work.

But Funes wasn't remembered as lucky. He was remembered as tragic. Perfect recall without selection isn't intelligence. It's just a very organized archive of noise.

The question for AI memory tools isn't how much to store. It's what kind of structure makes the stored things findable, useful, and — yes — worth discarding everything else for.

---

Asgeir Albretsen is the founder of Harbor.