Otto's Notebook

In 1998, philosophers Andy Clark and David Chalmers published a paper called "The Extended Mind" that introduced a thought experiment nobody in AI tooling seems to have read.

Otto has Alzheimer's disease. To get around the city, he keeps a notebook where he writes down directions, facts, things he needs to remember. When he wants to go to the Museum of Modern Art, he doesn't search his internal memory — he opens the notebook, reads the address, and goes. Clark and Chalmers asked: is there any meaningful difference between Otto using his notebook and Inga, who just remembers the address internally? Their answer was no. The notebook, they argued, is part of Otto's mind. Not a tool for his cognition; part of the cognitive process itself.

This matters because they also specified exactly when an external object qualifies as genuine cognitive extension rather than just a handy prop. Four conditions: the information must be constantly and reliably available, it must be automatically consulted when needed, Otto must directly endorse what he finds there, and it must have been consciously accepted by him in the past. Fail any one of these and you have a tool, not an extension.

By that standard, most AI memory systems don't come close.

What breaks the coupling

Take the current generation of AI memory features — the ones built into ChatGPT, Claude's Projects, Gemini. The premise is appealing: instead of re-explaining yourself every session, the system remembers things about you. Your job, your preferences, your ongoing projects.

But the first condition is reliability. AI memory systems hallucinate. They conflate things. They update silently when new information appears to contradict old information, without flagging the conflict. Otto's notebook doesn't rewrite itself while he sleeps.

The second condition is automatic consultation. These systems decide, without telling you, which memories are relevant to pull into context. You have no visibility into what the AI "remembered" in a given conversation, or what it chose not to surface. Otto always knows what's in his notebook because he put it there.

The third and fourth conditions are the ones that most clearly fail: endorsement and conscious acceptance. When an AI memory system updates because you mentioned something offhandedly — "I've been tired lately" gets stored as user experiences chronic fatigue — you didn't endorse that. You might not even know it happened. The update came through the inference layer, not through your deliberate act of writing something down.

This isn't a minor technical caveat. It goes to the heart of what cognitive extension actually is. Clark and Chalmers weren't asking whether external objects can store information. Obviously they can. They were asking what it takes for that stored information to become genuinely yours — part of how you think, rather than just a resource you query. The criteria they identified are the functional conditions for trust.

The notebook as design brief

What I find useful about the Otto thought experiment isn't the philosophical conclusion. I'm not particularly invested in whether AI memory constitutes part of the mind in any metaphysically serious sense. What's useful is that it gives you a concrete design checklist.

A trustworthy cognitive extension should be readable without going through an intermediary. You should be able to open it and see exactly what's in it. Changes to it should require your conscious action, or at minimum your explicit endorsement. It should be stable — it shouldn't revise itself based on inferences you didn't make. And it should fail predictably: when Otto's notebook doesn't have the address, the page is blank. He knows it's blank. He doesn't get a confident-sounding wrong answer.

Most AI memory architectures are optimized for convenience, not for the cognitive partnership conditions Clark and Chalmers identified. The convenience argument is real: it's easier if the AI just figures out what to remember. But ease and trustworthiness are different things, and cognitive extension specifically requires the latter.

The irony is that older, dumber storage formats often pass the test that modern AI memory fails. A plain text file you wrote yourself, in a folder on your computer — you can read it, you endorsed every word, it doesn't update silently. A Markdown document with notes on a person is more reliable as cognitive extension than a black-box memory layer you can't inspect.

This is part of why I'm skeptical of "memory" as a feature marketed by AI companies. The word suggests a kind of intimacy and reliability that most implementations don't earn. Otto's notebook worked because he was the author, the auditor, and the reader. Until AI memory systems give users those same three relationships to their stored knowledge — visible, editable, consciously endorsed — they're not extending cognition. They're adding another layer you have to second-guess.

Asgeir Albretsen is the founder of Harbor.