Most people's first AI assistant forgets them by Tuesday.
You open a chat box, explain your project, get something genuinely useful, close the tab. Next week you explain the project again. The thing that was sold as changing your life turns out to be a very articulate stranger you re-introduce yourself to every morning.
This is the other road. It starts on a Windows machine with nothing on it and ends with something that remembers what you told it last week, works on your actual files, and can be told what it must never do.
Six stops, in order. Every one has a full walkthrough on our setup pages; this is the map that tells you why each is there and what happens if you skip it.
What it costs
$20 a month. The gate is at sign-in rather than at install, so it is worth knowing before you start.
The terminal tool this road is built on — Claude Code — needs a paid plan. The free tier does not include it. Pro is $20 a month, or $200 for the year. Larger plans start at $100 a month. Those were the prices on the vendor's own page on 30 August 2026; check before you commit, because pricing moves.
Everything else here is either already in Windows or a free download.
Stop one: what the machine actually needs
Less than you have been led to expect, and this is the good news.
Windows 10 version 1809 (from 2018) or newer, 4 GB of RAM, an internet connection, and PowerShell or CMD — both of which came with your computer. That is the list.
You do not need Node.js, WSL, or administrator rights. If those mean nothing to you, that is rather the point — other guides ask for them and this road does not. The PowerShell already on your machine is the right one; you do not need to install a newer version first.
We checked that rather than assuming it, on a Windows machine here where this is already running.
Git for Windows is optional — our own setup pages used to say otherwise, and we corrected them. Install it and the assistant runs shell commands through Git Bash; skip it and it uses its own tool instead.
Stop two: the install
Open PowerShell and run one line:
irm https://claude.ai/install.ps1 | iex
If your prompt reads C:\ with no PS in
front of it, you are in CMD rather than PowerShell, and the line is
different:
curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd
That is the whole install. It keeps itself up to date in the background afterwards, so this is the only time you do it.
Stop three: the two-minute wall
The install finishes, you type the command to start it, and Windows tells you it is not recognised.
Close the terminal and open a new one. That is the fix.
Nothing is broken. The installer added a folder to your computer's list of places to look for programs, and the window you are standing in was opened before that happened — so it is still working from the old list. A new window reads the new one.
If it still isn't recognised after that, the folder has to be added
to that list by hand. It is %USERPROFILE%\.local\bin,
and our setup pages walk through where to put
it. Open a new terminal again afterwards.
You have not done anything wrong. This is the most common first stumble there is, which is why it has a heading of its own rather than a footnote.
Stop four: signing in
Run claude. It opens your browser, you sign in there,
and it stays signed in afterwards.
Then confirm the install is real:
claude --version
A working install prints a version number. We are deliberately not telling you which one to expect — it updates itself, so any figure printed here would be wrong within the week. A number is the pass; the number itself does not matter.
There is also claude doctor, which prints diagnostics
without starting a session. Run it the first time something feels
off, rather than guessing.
Stop five: the memory
This is the step that separates the two roads. It is also the shortest one here.
Two things carry knowledge between sessions, and they are not the same.
A file you write, called CLAUDE.md. Plain text,
sitting in your project folder, loaded at the start of every session.
Build commands, how the project is laid out, the corrections you are
tired of repeating. The rule of thumb is a good one: write down
anything you have now explained twice.
Notes it writes itself. This is on by default. It keeps what
it learns about your preferences and your corrections, in plain text
you can open, edit or delete. Typing /memory shows you
what it has.
Where it goes matters, and this is the step people get
wrong. Your project folder is just the folder holding whatever you
are working on — a website, a set of documents, your books. Put
CLAUDE.md inside it, and start the tool from inside it
too, by opening a terminal in that folder rather than wherever one
happens to be. Run it from somewhere else and the file is never read,
which looks exactly like the memory not working. Our
setup pages show that step for anyone who has not
done it before.
Keep the file short. The vendor's own guidance is to target under 200 lines, because longer files "consume more context and reduce adherence" — a long instruction file makes an assistant less reliable, not more obedient.
One thing worth knowing for later: a memory file is context, not a lock. If you need something that holds no matter what — a folder it must never touch — that is a hook, and both our setup pages and the official documentation cover how to write one.
Stop six: proving it works
The real test is not a command. Close everything, come back tomorrow, and ask it something only yesterday's session would know.
If it answers, you have an assistant. If it asks you to explain the
project again, your memory file is not being loaded, and typing
/context will show you which files actually made it
in.
What we could not test
Two steps, stated plainly: the install running to completion on a genuinely blank machine, and the first browser sign-in. Both are documented; neither was tested here, because we do not have a clean Windows machine. Every Windows install we can reach already has this working.
So the prerequisites, the update behaviour and the wall at stop three were measured on a real machine. The two steps above come from the vendor's current documentation, checked on 30 August 2026.
We would rather tell you which step we could not verify than let you find out on your own time.
What to do
Check your plan before you install anything. The free tier does not include this, and that is the wall people actually hit.
When the command is not recognised, close the terminal and open a new one. Do not start debugging.
And do not skip stop five. It is the only part of this that turns a chat box into something that knows you, and it is mostly writing down things you are already tired of saying.
The step-by-step, with every command, is on our setup pages.