Project Finch
An evolving cognitive architecture built around long-term learning. Finch asks what happens if an AI can improve continuously at your work over months and years — instead of starting over every conversation.
Today's models forget the moment a chat ends. Finch explores whether the right architecture — not just more compute — lets a single system remember, practice, and keep getting better at the things you care about.
FINCH, in full.
A friendly name on the surface — and an acronym that describes the architecture exactly: Federated Intelligence Node Cluster Holarchy.
Federated
No single point of authority. Each node owns and contributes to a shared substrate, so intelligence is distributed rather than centralized.
Intelligence
Every node is a complete AI in its own right — able to reason, learn, and operate on its own.
Node
The unit of the system — a Mac, a worker machine, a cloud teacher tier — each a self-contained participant.
Cluster
The nodes coordinate, sharing scheduled and event-driven work, so the whole accomplishes more than any single part.
Holarchy
In Arthur Koestler's sense: each node is at once a whole and a part — autonomous on its own, yet a component of something larger. Hierarchical but mutual — one drives, one offloads, one teaches, and each can still stand alone.
Put together, Finch is a federation of intelligent nodes that cluster into a holarchy — a system where every part is also a whole, and the shared substrate they build is the durable mind that outlives any single machine or model.
Persistent memory
Knowledge accumulates across interactions instead of resetting every conversation, so the system carries what it has learned forward.
Mastery loops
Staged learning cycles with verification gates — Finch practices, checks itself, and only advances once it has genuinely improved.
Failure-driven refinement
Errors aren't discarded; they become training material, turning mistakes into the next round of capability.
Transfer & ontology
Knowledge is organized so that what's learned in one domain can carry over and compound across others.
Longitudinal development
Capability is measured over months and years, not single sessions — the question is whether a system can keep getting better.
Local-first cognition
It runs on consumer hardware. Personal data and learned memory stay on the device rather than in someone else's cloud.
Learning in the open.
Finch is an experimental research project. Rather than hide the work behind a closed beta, we publish its progress as it happens — anyone can follow along at projectfinch.com.
The site streams a live system-health dashboard, a running devlog, mastery-progression tracking, and benchmarks, so you can watch the architecture develop day by day. There's no public chat endpoint yet — it runs locally on private hardware.
Follow Finch as it learns.
projectfinch.com →