What Forgetting Is For
Struggle and retrieval practice produce lasting knowledge. Re-exposure doesn't. What this means when you're building AI tools optimized for frictionless recall.
In 2006, Henry Roediger and Jeffrey Karpicke ran an experiment that should have changed how we build knowledge tools. Students read prose passages, then split into two groups: restudy the material, or try to recall it from memory without notes — struggling, gapping, getting things wrong. After five minutes, the restudy group performed better. After one week, the recall group retained nearly twice as much. The hard path, with its failures and gaps, produced knowledge that lasted. The easy path produced knowledge that vanished.
Robert Bjork calls this a "desirable difficulty." Conditions that appear to impede learning often maximize it. The struggle isn't an obstacle. It's the mechanism. Which is an uncomfortable idea when you're building AI tools designed to make knowledge effortless.
The direction we're heading
In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner documented what they called the Google Effect: when people know they can look something up later, they encode it less deeply. They stop forming strong memories of the information itself and start forming memories of where to find it. The internet became a transactive memory system — which is mostly fine, until the service shuts down or the context evaporates and the information is gone but so is your understanding of it.
AI accelerates this. Where Google helped you find information you might otherwise have internalized, AI synthesizes it for you. You don't retrieve a source and read it; the agent reads it and summarizes. You don't search your notes; the agent searches them and briefs you. One analysis of AI-assisted learning found that using AI for summarization left people retaining 22% fewer concepts in long-term memory than those who processed the material themselves.
AI knowledge tools are clearly useful. But there's something worth naming about what you trade away when knowledge becomes maximally frictionless.
What the friction actually does
When you write a note, the act of deciding what to write is itself a form of processing. You're forcing yourself to separate what mattered from what didn't. You're choosing language precise enough that you'll be able to retrieve it later. That choice isn't just filing. It's understanding.
Tagging something as a decision rather than a note is a commitment — you're claiming something about what kind of thing this is. Writing a person record means consolidating everything you know about someone into a form you can check against what you learn later. The structure imposes judgment. Judgment takes effort. That effort is where knowledge gets encoded.
This is what spaced repetition systems figured out. Ebbinghaus mapped the forgetting curve in 1885, and the finding has been replicated consistently ever since: the parts of your knowledge that decay quickly are the parts you haven't retrieval-practiced enough. Which means they're the parts you haven't fully earned yet.
Perfect AI recall doesn't fix this. It hides it.
What it suggests about how to build
None of this is an argument against using AI with your knowledge base. An agent that can search your notes and surface relevant material while you're working is genuinely valuable — possibly the first real improvement to knowledge retrieval in years. The question is which direction the friction flows: where does the cognitive work happen, and where does it get offloaded?
The tools that serve you over time are probably the ones that do the retrieval work but require you to do the encoding work. That make it easy to ask questions later, but require enough care at capture time that what goes in has been processed, not just filed.
Structured capture is one version of this. Writing a note that Alice changed roles and is now leading partnerships at a company you might work with is a different cognitive act than pasting a raw message into a notes dump. The form demands judgment. You have to decide what kind of thing this is, and that's the moment where it becomes yours rather than just stored somewhere nearby.
Whether that specific act is the desirable difficulty that makes knowledge stick, I honestly can't say. Roediger and Karpicke were studying explicit recall of prose, not the amorphous relational and contextual knowledge that personal knowledge bases actually hold. The research doesn't translate directly.
But the direction it points seems worth taking seriously.
Bjork's finding has held up for sixty years across hundreds of studies. The hard path beats the easy path. Memory built through retrieval practice outlasts memory built through re-exposure. That result doesn't go away because we invented better tools. It's just harder to see when the struggle is the part that got removed.
Asgeir Albretsen is the founder of Harbor.