Phoebe AI phoebe-ai.deIndependent research · machine cognition
The study of minds is full of ideas we’ve barely tried on machines.
Phoebe AI is a research project that applies cognitive science to language models — over very long conversations, across many providers, on its own infrastructure. The question is simple and open-ended: what happens when you take how minds work seriously as engineering?
The conversation is the instrument.
A century of work on attention, memory, forgetting, association, and the gap between recognising and recalling has shaped how we understand minds — and only a fraction of it has been turned on today’s models. Phoebe treats the frontier models as instruments and asks what those cognitive lenses predict, then tests it.
Long horizons and many models are the method. Short prompts hide most of what’s interesting; behaviour over thousands of turns doesn’t. And running the same probe across many providers is how you separate a quirk of one model from a property of this kind of model.
What falls out along the way.
Examples, not the mission — the point is the practice that produces them. Both are ongoing work, stated as such, not finished papers.
Memory for extreme long-horizon chat
An experimental long-term memory that has held one conversation past 15,000 turns with associative recall — surfacing an early detail from a few oblique cues, not keyword search. It’s what makes the long-conversation method possible at all.
A hypothesis for lost-in-the-middle
The most recent thing to come out: a falsifiable account in which middle-of-context positions read out weakly because they’re being reused as scratch memory — a feature, not a bug. Literature-checked, with cheap discriminating probes.
Bought-in intelligence, self-built everything.
A single research harness over many providers, so a finding can be replicated — or falsified — the moment it appears.
Come think alongside it.
Reviewers, cognitive scientists, interpretability researchers, and anyone who ends up here on purpose: questions, collaboration, and a run at the open problems are all welcome.
The latest hypothesis has a discriminating test — inspecting activation norms at middle positions — that needs interpretability infrastructure this project doesn’t have. If you can probe the layers, it’s built to be tested.