# How the assistant works toward a goal, and how to edit CORTEX.md safely

In NeuralVault 2.0 a new vault’s CORTEX.md tells the assistant to own a goal you state and only come to you when it must; an existing CORTEX.md is never changed, and when you edit one the assistant reads at most 40,000 bytes and drops the end.

*https://perpetualtechnologies.co.uk/support/cortex-working-toward-the-goal-and-editing-it — last checked 2026-10-11*

## Steps

1. A new vault’s CORTEX.md has a section headed "Working toward the goal", placed after "How to work". It tells the assistant to own a goal you state until it is met, try another route before asking, ask only when truly stuck or when the action needs your yes (all questions in one message, each with its recommended answer), decide small things and say what it chose, and finish with a short report.
2. It also says a refusal is a boundary and not an obstacle. If the safety guard refuses a call or you decline an approval, the assistant must not reach the same result another way. And autonomy never widens what it may do: the safety rules, the permission mode and the guard still decide every call.
3. Only new vaults get that section. An existing CORTEX.md is never touched, so an older vault does not have it. To add it, copy the idea into your own CORTEX.md.
4. To read what is in force, open Settings, then Intelligence, then Standing instructions. NeuralVault uses the first non-blank file of CORTEX.md, CLAUDE.md and AGENTS.md at the vault root. That screen only reads it. Edit the file itself in the vault folder with any text editor.
5. Keep it short. The file is sent on every turn, so each line is paid for on every message. The assistant reads at most 40,000 bytes. Past that the end is cut and never reaches the assistant, and Settings shows a notice saying how many bytes are never read.
6. Because the end is the part that is cut, keep the safety rules at the top. The starter file puts them first for that reason. Do not delete them or move them below long sections.
7. Before a big edit, save a copy of the file. If the assistant starts ignoring a rule, check the size first.

## How to tell it worked

Standing instructions shows the file in force with your change in it and no notice about bytes that are never read.

## If it does not work

**The Standing instructions screen says part of the file is never read.**

The file is over 40,000 bytes. Shorten it, or move detail into a note the assistant can read on request. The last bytes are the ones dropped.

**Your edit has no effect.**

An earlier file in the order CORTEX.md, CLAUDE.md, AGENTS.md may be in force instead. The screen shows which one wins and which are shadowed.

**The assistant stops to ask more than you want.**

In an older vault the goal section is absent. Add the idea to your CORTEX.md. It still asks before anything that needs a yes, whatever the file says.

## Related guides

- [Ask the assistant to remember something, and where it is kept](https://perpetualtechnologies.co.uk/support/remember-that-and-cortex-md)
- [Permission modes: Plan, Manual and Default, and the safety guard](https://perpetualtechnologies.co.uk/support/permission-modes-and-the-safety-guard)
- [Approval cards: approve, decline or rewind a write](https://perpetualtechnologies.co.uk/support/approval-cards)

## Still stuck

Bug reports belong in the app, in Settings, About. Those carry no contact details, so they
get fixed rather than answered. To get a reply, email us instead: that route targets two
business days. https://perpetualtechnologies.co.uk/contact