The skill is the protocol, the ledger is the memory

I work with AI all day, across very different kinds of projects — research-heavy work where a wrong citation is a disaster, web builds where the design is the point, and long-running personal projects that mostly need to not be forgotten. For months I kept hitting the same wall: every conversation started smart and ended amnesiac. The work was good in the moment and invisible a week later.

The fix turned out not to be a better model, a better prompt, or a better app. It was an operating layer — four pieces that do the jobs a model fundamentally can't do for itself. I remember them with one line:

The skill is the protocol, the ledger is the memory, the trigger is the discipline, and the connector is the feed.

The skill is the protocol

Every procedure I repeat is written down as a readable rulebook the AI follows — not tribal knowledge, not "the way we did it last time," a file I can literally open and inspect. When the AI improvises a procedure twice, that's the signal to write it down. The test of a good protocol: someone else could audit it.

The ledger is the memory

AI memory is a marketing term; files in version control are memory. Every project's state — what's done, what's in flight, what's blocked and on whom — lives in structured files that get updated as part of the work, and committed. If it isn't in a ledger, it didn't happen. The killer feature isn't recall, it's audit: git remembers what we believed on any given date, which matters enormously when the work involves deadlines and agencies.

The trigger is the discipline

Nothing runs itself, and rot is invisible without a schedule. Every protocol has a defined cadence — a weekly audit that reconciles the ledgers against reality, an end-of-day wrap-up that writes the day's learning down before it evaporates. The insight that took me longest: a system without triggers isn't a system, it's a pile of good intentions with a README.

The connector is the feed

Ledgers that humans update by hand die within two weeks — manual logging is how systems die. So the ledgers update from feeds that already exist: the repos themselves, email, calendars. The human answers questions only a human can answer; everything checkable gets checked against a source.

---

One more rule came out of the first week of running this, and it might be the most important one: memory proposes, sources confirm. The model is allowed to suggest anything from recall — a case citation, a best practice, a config incantation — but nothing enters the record until the actual source has been fetched and read. The model's memory is a lead generator, never a witness.

None of this is complicated. That's the point. The complicated version is the one you build accidentally, one clever conversation at a time, and it has no protocols, no memory, no discipline, and no feed — just vibes and scroll-back.