How it works
A vault your team reviews, and a boot block that points into it.
Cortex has two halves. One is a repository of notes a person wrote and a person approved. The other is an MCP server that hands an agent the short version at boot and the one note it needs when it needs it.
The three moving parts
- 01 · Vault
Markdown in Git, with an owner
Each rule, decision or guide is a note. Indexes point at them, so a reader can walk from the root to the owning wing to the note. A change lands as a pull request, which is what gives a rule a reason and a history instead of a reputation.
- 02 · MCP server
A small boot, the rest on demand
At boot an agent gets a short block: which note owns which moment, and how to ask for it. Then it calls for a note, a resolved bundle for the repository it is in, or a search result. The vault is not pasted into the session.
- 03 · Capability record
What already worked, with the evidence
Verified work is written to a record an agent can query before it gives up: the route, the proof it ran, and the limits that route does not cover. See a reply from the map.
What Cortex is not
Being clear about this saves everyone a trial. Cortex is a governed context layer, not a model, not a framework and not a transcript store.
Not a chat history
It holds what a person decided and wrote down, not a summary an agent made of its own conversation.
Not a wrapper
Nothing patches your agent. It speaks MCP, so the agent calls it the way it calls any other server.
Not a lock-in
The vault is your repository of Markdown files. Stop paying us and the notes are still yours to read.
Questions
The ones we are asked first.
Is Cortex a memory layer?
No. Cortex holds the rules, decisions and know-how a person wrote down and reviewed, as Markdown notes in Git with an owner and a history. It does not watch your sessions and write its own summary of them.
What happens to our notes if we stop using Cortex?
They stay what they already are: Markdown files in your own Git repository. The engine reads them, it does not own them, so stopping leaves you with a folder of notes your team can still read.
Does our context leave the machine?
The MCP server runs where your agent runs and reads the vault from disk. It makes no call to us. This website sets no cookies, has no analytics, and loads nothing from another host but the form’s bot check.
Do we have to install hooks or wrappers?
No. Any MCP-capable agent can call the server, and the rules are notes rather than code. Agents are asked to write a receipt of what they loaded at boot, and a check flags a session that did not write one.
Why not put the rules in one big instructions file?
A single file is paid for in context on every run and still has to be obeyed. We measured the part that decides it: whether the short boot text names the note that governs the moment the agent is in.
Who reviews a change to the rules?
You do. A change lands as a pull request against the vault repository, so a rule has an owner, a reason and a history, and an agent can propose one without being able to merge it.