The Vault

The Codex holds your files. The Wiki holds what you read. The Vault adds one thing on top: a safe way to change those files. Information sits in layers, changes flow one direction, your decisions are written down and never erased, and the AI can propose a change but cannot make it until you tick a box. This is what keeps a fast-growing Memory from quietly overwriting the truth.

This is the heaviest chapter in Memory, and for a solo operator it is optional. You do not need it to get real value from the Codex or the Wiki. Skip it for now. Come back the first time two of your own files disagree and you cannot tell which one is true. That moment is the problem this chapter solves. Until you feel it, the machinery here is work you do not need.

Why this comes last in Memory

The Codex and the Wiki both let the AI write to your files. That is fine when the stakes are low. A wiki page can be wrong and you fix it on the next pass. But some files are not low-stakes. Who you serve, what you charge, the big calls you have made. If an AI session quietly rewrites one of those because it read one convincing note, you lose the thread of why the business is the way it is.

The failure this prevents has a name: drift. New information lands (a call, some research, a test result) and either nothing updates, or things update in different directions. One file says one thing, another says the opposite, and neither flags the other as out of date. You end up with two answers to the same question and no idea which one you meant.

The fix is two ideas working together. Layers that make information flow one way. And a gate where the AI proposes a change but you make the call. Together they turn Memory from “files the AI edits freely” into “files the AI suggests against but never overwrites.”

Two ideas: layers and the gate

Information tiers (changes flow one way)

A “tier” is just a layer, a label that says how settled a file is. Every file gets one, written in its frontmatter (the few lines of settings at the very top of a markdown file, sitting between two --- lines).

Four layers, top to bottom:

Raw in. Notes, research, test results. They land in an _inbox/ folder and are never edited, only added to.
Core. Your decisions log, who you serve, what you charge. The settled truth.
Working docs. Playbooks, outreach, per-product notes. These come from the core.
Actions. To-dos, this week’s plan. These come from the working docs.

The one rule that matters: a change flows down, never up. Something new lands in the inbox at the top. You decide whether it changes your core. If it does, you decide what working docs and actions that change touches in turn. The AI drafts that walk down the layers for you. You approve the parts that are right.

The discipline this buys you: the newest file is not automatically the true one. A note you wrote today is raw input, not settled truth, until you have walked it up through the gate and written it into your decisions log. That single rule stops today’s enthusiastic brain dump from silently overruling a call you reasoned your way into months ago.

The confirm gate (propose, approve, apply)

The confirm gate is the safety catch. It means the AI can propose a change to a core file, but nothing in that file moves until you approve it. Three steps:

The AI reads a new input and writes a proposed change-set to a _proposals/ folder. Each item has a before, an after, and a reason tied to the source. It touches no core file.
You read the proposal and tick only the items you approve. Anything you leave unticked is rejected by default.
A second command applies only the ticked items. For each one it makes the edit, logs a line to a changelog, and adds an entry to your decisions log if it changed something core.

That is the whole thing. The AI does the drafting. You keep the decision.

The gate does a quieter second job too. Every time you approve or reject a change, you record how you decide. Those ticks pile up into a record of your judgment, which is what you would need later to hand more of the small calls to the system. You automate the decisions the gate has watched you make a hundred times, not the ones you have never seen.

Append-only decisions

Your decisions log is the anchor. One rule: you add to it, you never erase from it. Newest on top, each entry dated, each saying what you decided and why. When a new decision replaces an old one, the new entry names the one it replaced. You do not delete the old one. The length is the point. It is the record of why the business is the way it is, and it is what stops you and the AI from re-arguing the same call six months later.

The commands

Four installable commands do the mechanical work. You can run the whole gate by hand, but these make it fast.

/vault-ingest proposes only. It takes a new input, drops the raw source in _inbox/, and writes a change-set to _proposals/ with a before, after, and reason for each item. It never edits a core file.
/vault-apply confirms. It applies only the items you ticked, logs each one, and adds a decisions-log entry for anything core. If you ticked nothing, it stops and says so.
/vault-lint audits, read-only. It reports broken links, stale dates, files missing a tier, and two core files that disagree without one marking the other out of date. It suggests fixes and changes nothing on its own.
/vault-prioritize re-ranks your backlog against your decisions, so the top of the list is the thing actually worth doing next.

Pitfalls (the “if X, then Y” guide)

Situation	What to do
You already have a pile of files with no tiers	Do not reorganize first. Add a tier line to each one where it sits, then let `/vault-lint` surface the conflicts. Reconcile them one at a time.
Two core files contradict each other	This is the gate’s first job. Pick the winner. Mark the loser with a short “replaced by” note pointing at the decisions-log entry. Never just delete the old one.
You are tempted to let the AI just make the change directly	That removes the whole safety. If a change is so obvious it needs no review, it takes two seconds to tick. Leave the gate in.
A proposal has 12 items and only 2 are right	Tick the 2, ignore the 10. Partial approval is normal, not the exception. The rejected items cost nothing. They were never applied.
Your decisions log is getting long	Good. It is append-only by design. If it gets hard to navigate, add a dated index at the top, but never prune it.
You skipped tiers because your business is simple	Then it installs in an afternoon and stays simple. Tiers cost nothing when you have few files, and save you the day you have many.

A worked version: Maya’s

Maya runs an online store, a small service business, and a content channel. She is not coding any of this. She kept the Vault as small as it can be.

She made two folders and one file. An _inbox/ where anything raw lands: a supplier email, a note from a customer call, a screenshot of a competitor’s pricing. A _proposals/ where the AI writes anything it thinks should change. And a decisions.md, append-only, where her real calls live: “Dropped the wholesale line, too much support for the margin.” “Moved shipping to flat rate.” Each with the date and the why.

Her rule is simple. Raw stuff goes in the inbox, and nothing else touches her core files by itself. When she wants to fold new information in, she asks the AI to propose the changes. It writes them out as a list, a before and after for each. She reads the list, ticks the two that are right, ignores the rest, and only then does anything reach decisions.md or her working docs.

The thing to notice: Maya never built a big system. She built one honest habit. Nothing changes her settled files until she has seen it and said yes. The first time a customer note and her pricing doc disagreed, she knew which one was true in ten seconds, because the true one was in decisions.md with a date on it.

You’ve got it when

Raw inputs land in _inbox/, and nothing edits your core files by itself
decisions.md exists, is append-only, and has a few real dated entries
Every file that matters declares a tier in its frontmatter
The AI proposes changes to _proposals/, and nothing core moves until you tick it
You have run one real input all the way through: propose, tick, apply, logged
You have rejected at least one proposed item, so you know the gate is a real gate and not a rubber stamp

If the first four are solid and the last two are thin, the structure is sound. The discipline comes with use.

The /vault-ingest, /vault-apply, /vault-lint, and /vault-prioritize commands are in the starter kit and listed with what each does in ../03-habits/01-skills-and-commands.md.

→ Next: Senses (../02-senses/01-claude-code-setup.md). Leave Memory and install the setup that lets the Cortex perceive and act on the world.