When ChatGPT, Claude, Codex or another AI saves to a workspace, it names the
note with a folder and a name, like `people/ana-diaz` or `companies/acme`.
The folder says what the note is about. GBrain tells every connected AI the
folders to use and how to keep one note for each subject, each time it
connects, so nobody has to write those rules into the AI or add anything to
the notes.

## The folders a connected AI uses

| Folder | What goes there |
| --- | --- |
| `people/` | One note per person. |
| `companies/` | One note per company or organization. |
| `projects/` | One note per thing being built or run. |
| `meetings/` | One note per meeting: who was there and what was decided. |
| `deals/` | One note per deal. |
| `notes/` | Anything that fits none of the above. |

A connected AI is told not to make up a new top-level folder. Something that
fits no folder goes in `notes/`. The workspace's own AI also files into the
folders [Memory](/docs/workspace/memory) lists, such as Concepts, Ideas and
Originals.

## One note for each subject

Before it writes a note about a person, company or project, the AI is told to
search for the one that exists, by name and by the other names the subject
goes by, and to update that note instead of starting a second. When something
new disagrees with a note, it changes that note and keeps both claims, with
their dates and sources.

These are instructions the AI follows, not a lock. Two notes for the same
person can still happen, for instance when two AIs spell a name differently.
Ask any connected AI to merge them, and say which name to keep.

## No frontmatter needed

A note needs no frontmatter: the block of `key: value` lines some notes open
with. GBrain reads what kind of note it is from the folder, so a note in
`people/` is a person. Its title comes from the first heading, or from its
name when it has none.

Two fields are read when they are there: `type:` overrides what the folder
says, and `title:` overrides the heading. Nothing else is needed. Fields made
up for routing, such as a scope or a category, are kept as written, but
nothing in GBrain reads them to decide where a note goes, so they add nothing.
A note in a folder GBrain does not know is kept, and filed as a concept.

## Links between notes

A link written as `[[people/ana-diaz]]` to a note that already exists connects
the two notes when the AI saves. Once a day, usually overnight, GBrain reads
what came in and adds what the text says: that a person works at a company, who attended a meeting,
the dated lines in each note's timeline. That is how GBrain answers who someone
knows at a company. [Which AI does what](/docs/workspace/what-runs-where)
has the rest of that daily upkeep.

Naming a person in a sentence does not make a note for them. For someone new,
the AI makes their note in `people/` first, then saves facts and links to it.
A fact saved on its own lands on a person's note only when that note exists.

## What to tell the AI

Nothing extra. The [standing lines](/docs/tools/assistants) make it check the
memory first and save as it goes, and GBrain gives it the folders when it
connects. To file a batch of material, say what it is about: *"File these
meeting notes. Update the people and companies in them, and make a note for
anyone new."* Then open **Memory** to read what it filed, and
[Activity](/docs/tools/activity) to see each save.