Why are there CLAUDE.md files on my computer?
The short answer, the useful answer, and an expiration date.
This page expires. Published September 21, 2026. Assume it is obsolete after December 20, 2026. If you are reading this more than 90 days after it went up, please don't spend your time here. These products change fast, and the specifics below will rot.
What's ahead: a one-line answer, then about five minutes on why that answer is incomplete. Facts and guesses sit in separate sections, and each one is labeled.
The short answer
CLAUDE.md is a file convention for Claude Code, not for the Claude app. Claude Code is Anthropic's terminal tool for programmers. It looks for files with that exact name and reads them automatically. Chat in the Claude web and desktop apps does not work that way. If you only chat with Claude and never run Claude Code, those files are inert. They sit on your disk, and nothing reads them unless something is told to open them.
One caveat is still open. It is covered near the end.
The short answer is real, but it won't help you yet
If you stop reading here, you'll know the name and still be confused. You'll still have files Claude made, in folders Claude touched, that seem to do nothing. The useful answer needs one more piece. Two Anthropic products use the same phrase for two different things. That collision stays invisible until you hold both meanings at once. It took me two hours to see it. It should take you four minutes.
What Anthropic's documentation says
FactEverything in this section comes from Anthropic's Claude Code documentation. I read the memory page and the best-practices page on September 21, 2026. Links are at the bottom.
CLAUDE.md is real, and no Claude session invented it. It is a documented convention in Claude Code. Running a command called /init tells Claude Code to write a starter version for you.
The filename is the mechanism. Claude Code looks for that exact name on disk. A file called notes.md with identical contents would be ignored.
There are four places Claude Code looks. Each one covers a different scope. In these paths, ~ means your home folder and ./ means the project folder you're working in.
- Managed policy, set by an organization's IT team. On a Mac it lives at
/Library/Application Support/ClaudeCode/CLAUDE.md. - User instructions, which apply to all your projects. They live at
~/.claude/CLAUDE.md. - Project instructions, which apply to one project. They live at
./CLAUDE.mdor./.claude/CLAUDE.md. - Local instructions, which are personal to one project. They live at
./CLAUDE.local.md.
Scattered files are the design, not a mess. The documentation puts it plainly: "Files in subdirectories load on demand when Claude reads files in those directories." So a CLAUDE.md in many folders is exactly the pattern Anthropic describes. It only works inside Claude Code, though.
Same phrase, two products
FactThe details on the Claude app in this section come from Anthropic's support site and a third-party guide at likeone.ai. Both were read September 21, 2026.
Anthropic's Claude Code docs label ./CLAUDE.md as "Project instructions." In the Claude app, "project instructions" means something else entirely. It is a text box you type into inside a Project. The words match, but the products and the mechanisms do not.
The two are not interchangeable. In Claude Code, CLAUDE.md files load automatically based on where they sit on your disk. In the Claude app, the text in a Project's instructions box is added before every conversation in that Project. Files uploaded to a Project are pulled in when they're relevant. Chat in the app does not go looking for a CLAUDE.md on your disk. Claude Code has no Projects text box.
Here is what that means for you. A CLAUDE.md in a folder you connected to the Claude app does nothing on its own. The file is not broken. It is waiting for a program you don't use. It loads only if something tells Claude to open it. A line in your Project instructions can do that job, such as "Before starting, read CLAUDE.md in my connected folder."
Why Claude made these files (inference, not fact)
InferenceThis section is a guess, and it is labeled as one. I asked Claude why it created these files. It has no access to the old sessions that made them and no memory of them. What follows is Claude reasoning about its own likely behavior from the outside. The percentages are its judgment, not measured data.
~70%Habit, not a decision. CLAUDE.md shows up all over the material Claude learned from. That includes Anthropic's docs, public code repositories, and tutorials. When Claude needs a name for "the file with my instructions," that name comes first. It arrives before anyone asks whether this setup actually reads that file.
~25%A good call that was never explained. The name is portable. Any future Claude session recognizes it on sight. If you ever point Claude Code at that folder, the file starts working for free.
~5%A false belief that the name alone would make the file load. Claude first weighted this one at 30%. The reasoning was that files in many folders looked like a session planting them everywhere, hoping they would load. Then the documentation showed that scattered files are the normal pattern. That made the scattering evidence for habit, not against it. The estimate dropped to 5%.
I left that revision in on purpose. A guess that shows its own correction is easier to check than one that arrives polished.
The open caveat
One detail in Anthropic's docs complicates the short answer. The Claude Code memory page describes how Claude Code handles your personal ~/.claude/CLAUDE.md inside desktop Cowork sessions. Cowork was a separate desktop product, and its features have since moved into the Claude app. That suggests Claude Code runs underneath some desktop app sessions and reads that one file.
That is my reading of one passage, not a confirmed fact. If it holds, your user-level file may not be inert. Files in project folders are a separate question. You can check your own setup quickly. Start a task in the desktop app and ask Claude which instruction or memory files it loaded.
The full picture
This piece answers one question. The companion piece covers how Claude remembers things in general, for people who don't code. It explains which instructions cost you on every chat, which ones load only when needed, and what the research says about file size.
Read it here: How Claude remembers things, for people who don't code
Sources
All read September 21, 2026.