farrarilabsBooking
TOKYO12 OCT 2026

One memory for the whole fleet

TL;DR
Far runs several AI agents. They ship code, build personal projects, scout research and run a trading system. Each one used to learn alone, so a lesson paid for in one repo stayed in that repo. Now they share one knowledge base. Every agent writes lessons into it and reads from it before starting anything new. It holds 939 lessons today. A model proposes them, a mechanical check approves them, and nothing is stored without a pointer back to its source.

01 /The setup

The agents are not chat windows.

One ships code. One builds personal projects. One scouts research. One runs a trading system.

At first each kept its own notes. A bug one agent had already paid for in production was invisible to the others.

The fix was not a bigger model.

It was plumbing. One store, shared by all of them.

One store, four agentslessons written in, lessons read out
One shared store in the middle holds 939 lessons. Four agents sit around it: code, personal projects, research and trading. Each writes lessons into the store and reads lessons out of it.

02 /What the store is

Plain on purpose.

  • Markdown files are the truth. One file per lesson, with a name, a description, a type, a claim type and an origin.
  • The search index is derived from the files. It can be deleted and rebuilt at any time.
  • The files live in a git repo. One commit per run, so any run can be reverted.
  • Raw material is kept as it came in. Each lesson sits beside it with a pointer back. A summary never replaces its source.

03 /The three stages

Ingest. No model involved. Raw records are stored word for word.

Distill. A model reads the raw records and proposes lessons. It proposes only. It never writes to the store.

Commit. No model involved. This step checks that each proposed lesson points at real raw material, stamps where it came from, and writes it.

Every lesson is marked as a fact or an interpretation. A fact with no source pointer is rejected at commit.

Unverified claims are not dropped. Most engineering knowledge is interpretation, so dropping them would empty the store. They are flagged instead.

04 /Today

The store spans 3 domains: engineer, investor, operator.

Lessons in the store939lessons
Domains3engineer · investor · operator
Curated docs65docs
Raw records, about4,000records
What the store holdsraw records at the base, lessons distilled from them
Raw records
4,000
Lessons
939
Curated docs
65
Domains
3
Raw recordsLessonsCurated docsDomainsone scale for all four bars · raw records are about 4,000

Every agent writes to it. Work lessons and personal lessons sit side by side, each tagged work or personal.

Every day, each agent searches the store before researching anything from scratch. If a teammate already learned it, the agent reads the lesson.

One example. An agent about to scrape a travel site found, in seconds, that a teammate had already measured the site's bot wall, the stealth plugin that gets past it, and a pricing quirk.

New research steps spent: 0.

05 /The rule behind it

An autonomous publisher in the fleet once produced a fluent, confident quote that nobody had said. It was made up.

That set one hard rule. An agent runs unattended only when it transcribes. It does not run unattended when it authors.

Anything that summarises or editorialises goes through the same split the store uses: one side proposes, a gate decides, and a human approves what gets saved as learned knowledge.

The store is only as good as its weakest gate. So the gate is mechanical.

Back to notes