Most AI forgets every conversation. Finch compounds.
Finch is an AI whose knowledge grows — accumulating memory, verified skills, and corrected mistakes into something that keeps getting better and outlasts any single model. We build it in the open, with gates designed to refuse anything it hasn't genuinely earned. Watch it grow, live.
Live telemetry, updated continuously. No demos, no cherry-picking — this is the actual system.
- Graph densityGood
- Avg degreeGood
- Knowledge volume47Knowledge volume47
- Knowledge density59Knowledge density59
- Graph quality—Graph quality—
- Cross-domain connectivity—Cross-domain connectivity—
- Domain breadth—Domain breadth—
- Mastery progression—Mastery progression—
- Failure recovery0Failure recovery0
Why this is a different bet.
Today's most capable AI is extraordinary — and it starts over every time. Each conversation begins from zero; nothing it figures out for you on Monday is still there on Friday. Scale has produced models that know more, not models that grow.
Finch is the other bet. Instead of a bigger frozen model, it builds a living body of knowledge that accumulates — and because what it learns isn't locked inside any one model, it can keep that knowledge as the technology underneath it changes.
The wager is that architecture — not raw scale — is what lets an AI actually improve at your work over months and years.
It's designed to run close to you — private, local-first — so the things it learns about your work stay yours.
Where this is going.
Finch is a research project today — but it's pointed at something concrete: an AI that runs on your own hardware, keeps getting better at the work you care about, and answers to you rather than a server farm.
It's the flagship project of Eidetikos, PBC — a public benefit corporation building AI held to long-term human benefit by charter. The company is where the mission, the research agenda, and the team live. This site is where you can watch the work happen, day by day.
- 2026-06-09
First topic to fully clear the expert gate
All six sub-areas on the writing topic crossed from proficient to expert in 48 hours — the first topic where every sub-area is at the expert tier. The dual-judge architecture — a primary judge paired with an independent adversary — caught real flaws on the way; the conservative aggregator pulled scores down where the adversary disagreed.
Read more - 2026-05-31
Graph health metrics live on the public site
Structural-quality signals — predicate vocabulary, leaf percentage, and cross-domain bridge entities — now stream to /graph alongside the size metrics. Substrate density is visible the same way substrate size is.
Read more - 2026-05-27
The relation axis stopped diverging
The predicate space had been bloating with near-synonyms — every new extraction was minting fresh relation names instead of reusing canonical ones. Canonical-first extraction shipped; most new triples now land on the curated vocabulary instead of inventing variants.
Read more
- 2026-05-23
Verify queue allocated all five slots to one domain
First overnight auto-verification run gave all five nightly slots to analysis because iteration order over unverified candidates favored it. Writing and python got nothing that night.
- 2026-05-19
Tutor session crashed mid-run
An analysis tutor session crashed about ninety minutes in. No useful output was captured before it failed — the verifier hit a problem state and stopped early. Logged for replay.
- 2026-05-17
Silent fallback masked the real fail
A retry layer was swallowing failures and returning fake-success values. Metrics looked healthy while the underlying job had silently dropped. Fix: failures propagate or get an explicit no-confidence flag.
Read more
What makes Finch different
- 01
Persistent memory
Every interaction becomes part of a growing substrate. Finch does not reset to a fresh context window each conversation.
→ Core conceptcontinuity - 02
Mastery loops
Topics advance through staged learning cycles: warmup, generation, verification, and multi-step tutor validation before promotion.
→ Core conceptearned progression - 03
Failure-driven refinement
Failures are preserved as learning material. Incorrect reasoning, weak retrieval paths, and failed attempts become future remediation targets instead of disappearing after the session ends.
→ Core conceptfailure becomes training signal - 04
Transfer and ontology
Concepts learned in one domain reinforce related domains while ontology normalization prevents the graph from fragmenting into disconnected synonyms and duplicate concepts.
→ Core conceptknowledge organization + transfer - 05
Longitudinal development
Capability is expected to compound over months and years through accumulated verification history, tutor cycles, and recursive correction — not isolated prompt sessions.
→ Core conceptdevelopment over time - 06
Local-first cognition
Finch is designed to run on standard consumer hardware with persistent local memory and low ongoing compute requirements.
→ Core conceptarchitecture over brute-force scale
- 2026-05-22
The first month — how Finch got here
Backdated devlog of the first 34 days, from an empty repo to a substrate doing real work. Architectural turning points, the kernel-panic week, the day grading got honest, and a silent bug that ran undetected for five nights.
Read - 2026-06-19
The second month — where Finch is now
Sixty days in. Long enough to have real numbers, short enough that the original idea is still recognizable. A retrospective on what changed, what didn't, what's still broken, and the three things that surprised me most about teaching a system to be honest.
Read
- 2026-06-11
Analysis is fully proficient — the first topic with a 16-sub-area surface to clear the rung
Two days ago I wrote about writing crossing the expert gate — six sub-areas, fully promoted. Today analysis crossed proficient on every one of its sixteen sub-areas. Same ladder rung as writing was on a week ago, but on a surface 2.7× the size. This post is about why the bigger surface matters more than the rung itself, and what the next gate looks like from here.
Read - 2026-06-09
The first topic to clear the expert gate — what changed over the weekend, and the night it actually mattered
Five days ago I asked whether anything would ever clear the expert gate. Over the weekend I built the machinery to find out. Over the next two days, six sub-areas on the same topic crossed it — the first topic to be fully expert across every one of its parts. This is the post about the work between those two moments, and the nightly that proved the machinery does more than promote on demand.
Read - 2026-06-04
The first proficient promotions — and what I added to the public dashboard to keep myself honest
Yesterday's post ended with "the system gets to make its case." Overnight, it made it. Nine sub-areas crossed from competent to proficient — the first promotions in 51 days. The story of why they happened now, what the audit found, and the new graph-health panels I shipped to the public site so the next promotion has nowhere to hide.
Read
A short digest when something real happens — promotions, milestones, the occasional honest setback. No spam, no account, unsubscribe anytime.
For press, partnerships, research, or investment inquiries, reach the team at Eidetikos — the public benefit corporation behind Finch.
Contact EidetikosNOTICE // What this site is not: a product page, a chat interface, or a claim of general intelligence. Finch runs locally on private hardware. This is the public log of what's inside — a project of Eidetikos, PBC.