Memory is the part that makes a workspace different from a chat window. What it learns is written down, filed, and still there next month.
It is not a hidden index. Open Memory and you are looking at the same notes the workspace reads when it answers you.
Why it is organized this way
GBrain files memory the way the open source GBrain project lays it out. Each rule is there to keep a memory from going stale.
One home for each thing. Every person, company, meeting and idea gets one note, in one folder. A fact lives in one place, so there is never a second copy saying something older. Someone who is a customer and a friend still has one note in People; links to other notes carry the rest.
What is true now on top, the history underneath. A note opens with a summary of where things stand, rewritten whenever something changes. Below a line is the dated record of how GBrain learned it, which is only ever added to. Read the top for the state of things and the bottom for what happened and where it came from.
The answer is already written. A plain search tool reads raw documents again for every question. GBrain does that work when something comes in: it updates the notes, links them, and flags where two sources disagree. A question is answered from notes that are already up to date.
Everything new updates what it touches. A meeting, an email or a conversation updates the notes of the people and companies in it, so the memory grows as a side effect of the work.
Under the notes, GBrain keeps one record for each person and company with every name and address they go by, each fact with where it came from, and the links between them. That is how GBrain answers “who do I know at this company?” without reading every note.
The folders
| Folder | What goes there |
|---|---|
| People | One page per person: who they are, how you know them, what they said. |
| Companies | One page per company or organization you deal with. |
| Meetings | One page per meeting: who was there and what was decided. |
| Projects | Things being built, with what is done and what is next. |
| Ideas | Possibilities nobody is building yet. |
| Concepts | Ways of thinking you would teach someone. |
| Originals | Your own thinking, in your words. |
| Sources | What came in whole: imports, transcripts and documents. |
A folder you have no use for stays empty. When something fits none of them, GBrain makes a new one.
How good it is
On LongMemEval, the standard test for long-term memory in AI, GBrain put the right conversation in its top five results for 95% of questions (449 of 470). The run, the settings and the scorecard are published in the open, including the twenty-one it missed.
The design is open too. It starts from Andrej Karpathy’s idea of an AI that keeps its own wiki and takes it from research notes to a working memory fed by your email, calendar and meetings. Read the design in the open source GBrain repository.
What is in there
Folders group notes by what they are about. Every memory starts with the folders above, and GBrain adds one when something fits none of them. You can ask it to reorganize, split, or rename.
Notes are the unit. One note is one thing worth remembering: a person, an idea, a decision, a document it read. Each is a plain file, so nothing is locked in a format only we can open.
How things get in
You tell it. Anything you say in chat that is worth keeping, it keeps. “Remember that Ana runs procurement” is enough.
It reads what you give it. Paste a document, drop a file, or point it at a page and it takes notes on what it read.
Skills write. A skill that researches a person or works through an inbox leaves its findings in memory, so the next run starts from the last one.
Scheduled work tidies. Overnight the workspace files what came in, links it to what it already knew, and cleans up after itself. That is why a workspace you have been feeding for a month answers better than one you set up yesterday.
Adding a document
The Memory page takes documents as well as chat does. Under the folders, press Add a document and pick one or more PDF, Markdown or text files. Each one becomes a page in Sources, and from then on every chat in the workspace can search it.
While GBrain reads a file, its row says Being read into the memory. Once the page is written, the row says since when and opens the page. A file GBrain cannot read says why: a scanned PDF, for one, has no text in it until it has been through text recognition.
Adding a file with the same name again replaces its page.
A row that is still being read, or could not be read, opens the document’s own page: its file, its size, and where it is. Remove there takes the document out, with its file and, if it was written, the page it became. It asks first, on a page of its own that lists what goes.
Reading the memory well
Ask it to look. “Search the memory and tell me what we decided about pricing” beats “what did we decide about pricing.” A casual question can get answered from general knowledge rather than the material, and it sounds equally confident either way.
Ask for the note. An answer it cannot source is an answer you should not repeat. It will name the note, and you can open it.
Correct it in chat. If a note is wrong, say so. It will fix the note, not just the answer, so the correction survives the conversation.
Reading the memory from somewhere else
Memory is not only for chat here. Claude, ChatGPT and OpenClaw can search it and add to it through one address, with you choosing whether they may write. Use the memory from other AIs.
What the memory does not do
Memory is per workspace, and everyone in that workspace sees all of it. There is no per person view inside one, and no way to hide part of a workspace from someone who has access to it. If a group should see a different set of material, give them their own copy.