System Online
----.--.-- --:--:-- UTC
PROJECT FINCH
Day 46 · since first fileREC / 00

An evolving cognitive architecture built around long-term learning.

Finch is an experimental AI that accumulates memory, verified skills, and corrective feedback over time. Instead of isolated conversations, the system is designed around continuity, adaptation, and cumulative capability growth.

▍The pulse
Daysince first file
46
2026.04.19 → today
ACTLast active
since last update
TSKIn queue
0
tasks pending
SYSSystem health
80/ 100
GOOD
  • Graph densityGood
▍Last 24hΔ
CALIBRATING // 24-hour deltas appear after the pusher has been running for a full day. The first snapshot is being written now; check back tomorrow.
▍The climbFull ladder →
No ladder data yet.
▍The question

Why not just use existing AI tools?

Frontier AI models are already excellent at general knowledge, one-off tasks, and even large projects like building apps. I use multiple model families daily. But most require enormous compute, large GPU clusters, or heavyweight local models that consume significant system resources — and access isn't cheap: capable subscriptions routinely run $100+ per month across the major providers. Even then, long-running projects can still lose continuity and context over time.

Finch explores a different question entirely:

What if an AI could improve continuously at your workflows over months and years instead of conversations?

What would that look like after a month? A year? Three years?

Persistent memory, mastery tracking, failure recovery, and long-term adaptation are the core experiment. The belief behind Finch is that architecture — not just scale or raw compute — may be the key to building AI that grows continuously inside one system, not across new releases.

▍The horizon

Would Finch ever become public?

That depends entirely on whether the architecture proves itself over time. But Finch has been designed from the beginning with the possibility of a future public release in mind — not just as a research project, but as something other people could eventually run, extend, and build on themselves.

If Finch succeeds at what it's attempting, the long-term plan is to release a public version others can use, study, and contribute to.

Recent breakthroughsWINS
  • 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

    Slot cadence now streams to the public site

    Finch's run scheduler now publishes its slot timing — currently-running, last-completed, and next-fire signals — so the public site reflects the real heartbeat instead of inferring it from staleness.

Recent failuresLOSSES
  • 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
▍Operating principles

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
▍From the devlogAll entries →
Pinned·Oldest first
  • 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
Recent·Newest first
  • 2026-06-03

    Cleanup week — how a routine refactor almost deleted 41% of the graph

    A week ago I shipped a normalization that worked exactly as designed. The morning after, the system tried to delete 41% of the graph. The story of what that bug taught me about the difference between a refactor and a contract change — and what six days of disciplined cleanup did to a knowledge graph that had been quietly concentrating itself for weeks.

    Read
  • 2026-05-28

    The snowflake, 17-days stuck, and what it took to see both

    Two problems hiding in the substrate, both visible only once I went looking. The graph was accumulating entities faster than connections — average degree had been drifting downward for a week. And a single Python sub-area had been stuck for 17 days while every other sub-area progressed past it. Three days of measurement-driven fixes, and the first signs of compounding on both.

    Read
  • 2026-05-25

    Analysis competent, marketing online, the reliability stitch

    Analysis just promoted its last sub-area, making it the second domain fully at competent after writing. Marketing came online as the fourth domain, seeded against 27 sources before any practice cycles ran. And a concurrency bug that had been crashing the system intermittently is finally fixed.

    Read
▍Get the next post in your inbox

New journal entries delivered when they publish. No spam. Unsubscribe with one click.

▍Get in touch

Questions, feedback, or just curious — drop a note. Replies come from a real person, not a chatbot.

8–4000 characters

NOTICE // 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.