Glasp Research
What millions of highlights reveal about how people read — and how AI agents handle inherited memory. Every paper ships with its reproducibility bundle.
A small reader-trained model still beats October 2026's frontier models — and more reader pages help it where a stronger tie-breaking teacher does not.
Where a reader is on the page predicts their next highlight better than which readers they resemble — 47% to 29% in the top five.
When agents act on notes, plans and memories written earlier, the evidence pins down the shape of the problem better than its size.
How a memory is worded decides which record an agent goes back and checks — and appending an id can cancel the wording that worked.
Handed a stale memory, agents act on it three times out of four — and the error is allocation, not reasoning.
A provenance link that is available is not a link that gets read.
Machines can now recover 60% of the predictable part of crowd attention.
Language models agree with each other more than any of them agrees with human readers.
Scored fairly, a single human reader still edges out the best language model.
Reading identity is a durable trait: a six-month snapshot predicts you years later.
What a crowd will highlight on a brand-new page is predictable, and the win is in the long tail.
Readers form camps inside a document, but the camps do not travel with them.
Personalization lives in what you choose to read, not in where you highlight.
Within a page, people mostly highlight the same things.
Measuring AEO honestly means separating it from the platform's own growth.
Most research on highlighting comes from small lab studies. Glasp's platform captures how people highlight in the wild, at scale, across years. Publishing what we find, including the negative results, keeps our product decisions honest. Questions or collaboration ideas: [email protected].