What Lab Notebooks Got Right
Scientists spent centuries figuring out what makes a record trustworthy. Personal knowledge tools mostly ignored the answer.
Faraday repeated experiments he'd already done. More than once. He'd complete a line of investigation, publish the results, and then — sometimes months later — return to the same setup and run the same experiments from scratch, apparently with no memory of having done them at all.
He knew this about himself. So he built a system that didn't depend on his memory being reliable. Between 1832 and 1860, he kept a continuous record of his laboratory work, numbering each entry sequentially from 1 to 16,041. Every entry was dated. Failures were documented alongside successes — in fact, failures were considered important to record precisely because they wouldn't make it into published papers. The record was kept not just for Faraday's own reference, but as a trust artifact: something a future reader could consult to understand not just what he found, but how and when and under what conditions.
His notebooks are now on the UNESCO Memory of the World Register. They've been used to settle patent disputes. They survived him by more than a century and are still being analyzed by historians of science. None of this would be true if he'd just used the Victorian equivalent of Apple Notes.
What the conventions were actually for
The rules that evolved around lab notebooks look like bureaucracy from the outside. Bound pages rather than loose-leaf, so nothing could be inserted retroactively. Ink rather than pencil, so entries couldn't be altered without visible damage. Dates on every entry. Signatures. Witnesses who weren't co-inventors, who could confirm that what was recorded matched what actually happened. No blank spaces left between entries: unused lines were struck through to prevent backdating.
These conventions weren't developed because scientists are fussy. They developed because notebooks had to hold up in court. Patent disputes turn on prior art — who knew what, and when. FDA regulatory audits turn on the chain of custody for experimental data. A notebook without these properties isn't just poorly organized. It's inadmissible.
What the conventions actually encode is a very specific idea: that a record should be trustworthy to someone who wasn't there. The witness, the date, the sealed binding — these exist to let a future reader answer the question "can I trust this?" without having to trust the person who wrote it.
Most personal note-taking tools have never thought about this problem.
What we optimized for instead
Digital note-taking arrived and solved different problems. Search got better. Sync got faster. Tags proliferated. But the basic model of a note remained a blob of text with a creation timestamp buried in metadata, visible to nobody.
There are no witnesses in Notion. Obsidian doesn't distinguish between what you wrote last year and what you quietly replaced it with this morning. Most apps make it functionally difficult to record failure — the UX is designed for capture of the interesting, not for systematic documentation of what didn't work.
The culture of recording failures disappeared entirely. Faraday documented experimental failures because he understood that a record containing only successes is a curated narrative, not an honest account. Published papers were already doing that job. The notebook existed to capture what didn't make it in.
If your notes app only lets you record things you think are worth keeping, you're building a curated narrative. Which is fine, until you need to retrace your reasoning six months later. Or until someone else has been editing your knowledge base and you want to know exactly what changed.
What AI makes urgent
When you're the only one writing to your knowledge base, the trust problem is manageable. You know roughly what you meant. You remember the context, at least partially. The gaps are annoying but survivable.
When an AI writes to your knowledge base, none of that applies. The model doesn't remember. It can't be asked what it was thinking. It won't be there to provide context when you read the entry two years later. And unlike Faraday's forgetfulness, which was bounded by his humanness, an AI agent can generate and update entries faster than you can inspect them.
The lab notebook conventions solve this problem directly. Every AI write needs to be attributable, timestamped, and separable from what you wrote yourself. AI proposals need to be entries in their own right — something you review before they get accepted into the main record. The audit trail isn't an afterthought. It's how a future reader, including your future self, can understand not just what your knowledge base contains but how it got that way.
This is not a new idea. It is a very old one.
Scientists figured it out because they had to. Patent courts and regulatory agencies forced the question: what does it mean for a record to be trustworthy? The answer they arrived at — attestation, provenance, continuous chain of custody, documentation of failures alongside successes — applies just as well to a personal knowledge base being written to by an AI agent.
The irony is that we spent thirty years building note-taking tools and almost none of them inherited these properties. The lab notebook was sitting there the whole time.
Asgeir Albretsen is the founder of Harbor.