How to organize AI chats into projects on your Mac
- A project is a durable container: files, context and chats that belong to one piece of work.
- It means not pasting the same background into every new conversation.
- Point a project at a codebase and the AI can find the relevant file instead of guessing.
- Everything stays on your machine — a project is local state, not a cloud workspace.
The hidden tax on chat assistants is re-explanation. Every new conversation starts from nothing, so you paste the same context again, and the quality of the answer depends on how much of it you remembered to include.
What a project changes
Instead of context living in your head and being re-typed, it lives with the work. Open a chat inside a project and the assistant already has the background: the files, the constraints, the prior decisions.
For code the difference is sharper still. With a codebase attached, the assistant can locate the function you're describing rather than asking you to paste it — which is the difference between a colleague who has seen the repo and one who has not.
What actually gets indexed, and why it depends on your Mac
“Point it at a codebase” raises a fair question: all of it? The answer is that the index is sized to the machine, in five bands, because an index that fits a 128 GB Studio would swamp an 8 GB Air. These are the limits the shipping code applies.
| Your Mac’s RAM | Files indexed | Directories | Time allowed to walk the tree | Largest file read |
|---|---|---|---|---|
| Under 16 GB | 3,000 | 1,000 | 45 s | ~1 MB |
| 16–32 GB | 10,000 | 3,000 | 90 s | ~2 MB |
| 32–64 GB | 25,000 | 8,000 | 180 s | ~3 MB |
| 64–128 GB | 50,000 | 15,000 | 300 s | ~4 MB |
| 128 GB and up | 200,000 | 60,000 | 600 s | ~16 MB |
For scale: a real working repository — the one Outlier itself is built from — indexes at about 3,000 files, which fits inside the smallest band. Most projects are not near these numbers, and the bands exist for the monorepo that is.
The part worth knowing is what happens when a project does exceed them. It is not silent. The index records that it stopped at the limit so the interface can say so, and the file tree the assistant sees is paths-only and therefore far larger than the content budget — so on a normal repository it sees the whole tree even though it has read only part of the contents. When something is omitted, the assistant is told it can still open or search that path on request. Nothing is hidden; some of it is merely not pre-read. That distinction is the difference between an assistant that says “I cannot see all the code” and one that quietly answers as though it had.
codebase.py, the RAM-tiered index limits): files
3,000 / 10,000 / 25,000 / 50,000 / 200,000 and directories 1,000 / 3,000 / 8,000 / 15,000 /
60,000 across the bands under 16, 16–32, 32–64, 64–128 and 128+ GB;
tree-walk budgets 45/90/180/300/600 s; per-file read caps ~1/2/3/4/16 MB. Truncation
is reported rather than silent. These are enforced values, not intentions.Setting one up
- Create a project for a real unit of work, not a topic.
- Attach the files or codebase that define it.
- Add standing context — the conventions and constraints you'd otherwise repeat.
- Chat inside it, and stop re-explaining.
Getting more out of it
- Narrow beats broad. A project scoped to one codebase outperforms one scoped to "work".
- Write the constraints down. The things you'd tell a new teammate are exactly what belongs in standing context.
- Projects and memory are different. Memory is what the AI knows about you across everything; a project is what it knows about this work.
Common questions
What is an AI project folder?
A container that holds files, standing context and the chats for one piece of work, so every conversation starts with the background already loaded.
Can a local AI read my whole codebase?
It indexes the codebase you attach so it can retrieve the relevant parts on demand. Everything stays on disk.
How is a project different from memory?
Memory is cross-cutting knowledge about you. A project is scoped context about one body of work. They complement each other.
Is project context uploaded anywhere?
No. Projects are local state on your Mac.
Try it on your own Mac
One signed, notarized download. No account, no token bill, and it keeps working with the Wi-Fi off.
Free: Nano + Lite. Pro: $249 once · or 4 × $62.25; $124.50 for the first 25 · or 4 × $31.13. macOS 26+. In the US, Klarna or Afterpay at checkout: four payments, two weeks apart. Refund window: 30 days, no questions asked.
Download free (290 MB) Buy ProOn your phone?