Notes from the lab
Personal Models: how we are designing an intelligence that learns one person
The design behind Anjo's next research program: what a Personal Model is, what it may learn from, how it is deleted, and what would prove us wrong.
Why we withdrew our memory benchmark number
We published a LongMemEval-S retrieval score, found the harness could count a miss as a hit, and withdrew it. The re-run held, and the figure stays in the repo.
Why we filter 'last time you said' out of Anjo's replies
A language model can fabricate a shared memory in one sentence. Anjo removes unsupported ones before display. How the guard works, and where it fails.
Measuring whether conversations leave you better, against our own engagement
Anjo tracks whether a conversation helped, not how long it ran. The wellbeing trend, the outcome ledger, and the rule that self-report alone moves nothing.