The house where your AI’s context lives.

Everything an AI has in front of it when it answers: your question, the conversation so far, and any files or instructions attached to it. That is all it knows in that moment, and a new chat starts with none of it. ContextHouse is the folder you attach so that your week (Now.md), your projects, your decisions and your people are in front of the AI every time, without you retyping them. Two of the 55 files, CLAUDE.md and AGENTS.md, work the same way in the other direction: they brief the AI on the house rules before it touches anything.

Built-in memory lives inside one company’s app: a summary it extracted from your chats, which you can view and delete but not really organize, and which moves to another company’s app as a copy-paste job at best. The house keeps that context as plain files instead, so you can read them without an AI and hand them to a different one. It has rooms because one folder with everything in a pile helps nobody, and a steward because every filing system drifts unless someone checks it weekly. The AI works in those files as a guest, under rules you can read in CLAUDE.md.

Yes, that idea, done as a house for your AI’s context. “Second brain” is the name people use for keeping your memory outside your head in a linked system, from Luhmann’s paper cabinet in 1981 to today’s note vaults. ContextHouse is that idea with the setup already done and an AI reading the files. If the term is new to you, or you want the honest case for keeping one, it’s on Why ContextHouse.

In practice: Claude or ChatGPT on a paid plan, using their desktop apps. More generally, any assistant that can read and write files in a folder works; the instruction files ship in both standard conventions (the two instruction formats most file-capable AI tools read), so your assistant is briefed the moment it opens the folder. Which tools do that fully, which partially, and what they cost shifts every few months, so the as-of-dated comparison ships inside the house (System/Model_Recommendations.md) and nudges you toward a good setup on day one. And when you want the current answer, ask the assistant you already have to look it up; that question is one any model answers well. Fair warning: these tools carry their own subscription or usage cost, separate from this purchase; budget roughly $20/month as of September 2026. With a plain chat AI you drive by hand: paste CLAUDE.md at the start of a session and it knows the rules and coaches the rituals, but you move the text in and out of files yourself. Good preview, not the working system. The one thing that doesn’t work is running it with no AI at all, since operating the files is the assistant’s whole job here.

For a look around, yes; for the working system, no. The browser chat cannot read and write files in a folder on your computer, and operating the files is the assistant’s whole job here. The fix is one step, once: install the free ChatGPT desktop app, sign in with the plan you already pay for, and point it at the folder. About ten minutes, and the tool guide inside the package (System/Model_Recommendations.md) walks the exact clicks for each AI. Your subscription and login stay the same; only the window changes.

In conversation: yes, fully. You talk to your AI in your language and it answers in yours. The guided setup runs the same way: say “hilf mir, das einzurichten” and the whole move-in happens in German, and what you tell the house lands in the files in your words, in your language. What stays English: the folder’s structure (file names, headings, the instruction files your AI follows), the handbook, and this site. That layer is English because it is what AI assistants follow most reliably; you read it once, your AI reads it every day. If you can read this page, you have all the English the house needs.

One download: a zip containing the ContextHouse folder, 55 plain-text files (54 Markdown files and one illustrated handbook for your browser). Unzip it wherever your files live; a synced folder like Drive or OneDrive is perfect. Inside: the rules of the house, twelve ready workflows, the worked example, the manual, and the AI instruction files in both standard conventions, so your assistant is briefed the moment it opens the folder. No app, no account, no installer.

After unzipping, you point your AI at the folder and say “help me set this up”. The house then runs its own move-in: a first session of about fifteen minutes where it explains the rooms, asks about your life, and writes the first pages with you, in your words. The rest happens in short optional sessions over the coming weeks, at your own pace. No questionnaires and no homework between sessions, and progress is remembered in a file, so any session with any AI picks up where the last one stopped. You can also skip all of it and furnish the rooms by hand in about twenty minutes; nothing in the house depends on the guided setup.

In the desktop apps, yes: a new chat starts by choosing the folder, the same menu as in the picture above. Claude Code users simply open a terminal in the folder; it reads the house rules on its own. If you forget, nothing breaks: the AI answers like a normal chat that day, the files stay exactly as they were, and the next connected session picks up the written state.

The files live on your machine (or your own Drive/OneDrive). Nothing about this product phones home; there’s no app, no account, no server behind it. But one sentence must be said plainly: when your AI reads a file, that file’s content is sent to the AI company (Anthropic, OpenAI, Google, whichever runs your assistant). Reading is sending; that is how these tools work. What the AI never reads never reaches the AI company. And truly private material (ID scans, medical records, password lists) relies on no rule at all: it lives outside the folder your AI is connected to, out of its reach entirely. A folder the AI can’t reach beats any instruction it might follow. The full plain-language picture, what travels, when, and what stays outside, is System/What_The_AI_Sees.md. Read it once; it takes two minutes.

Good free starting points exist, including well-starred ones, and if you enjoy building systems you should try one. What they leave with you is the judgment work: deciding the structure, writing the rules, inventing the workflows, and keeping the whole thing honest six months in. ContextHouse arrives with that work done and tested in daily use, and adds what the free kits skip: rules built for safe AI collaboration (everything the AI writes is visibly marked, and the AI is never allowed to declare its own work trustworthy), and a maintenance loop that never runs unattended: a fifteen-minute weekly check where every change is a list you approve, one system change per week. On top of that, the parts that make day one easy: a guided setup that interviews you instead of handing you homework, a worked example so you see the loop before you commit to anything, and a human on the support address. If a free kit gets you there, take it. This is for the person who wants to move in, not build the house.

No, they get along. Obsidian is an editor: a good way for a human to read and write Markdown. ContextHouse is 55 plain-text files, so you can open the folder in Obsidian today and browse it there for the rest of its life; nothing needs converting. What Obsidian doesn’t ship is what you’re buying here: the instruction files that brief your AI, the rules for safe AI collaboration, twelve ready workflows, and the weekly check that keeps the system honest. Build all of that inside Obsidian yourself and you’re back at the lost weekend. Drop the house in, keep Obsidian as your reading chair while Claude does the filing.

A template gives you pages inside one app. This gives your AI an operating manual, in plain files no app owns: rules it follows, workflows it runs, memory it maintains. A Notion template can’t tell your AI what it may touch, and it can’t run a weekly check on itself; both are the point here.

Yes, gradually, and without a big import day. Anything can go into the inbox as it is (files, exports, stray text), and the sweep proposes where each piece belongs; your old notes get filed the same way new ones do. If you already keep an Obsidian vault, the files are plain Markdown on both sides, so you can copy folders across and let the AI help sort them room by room. There is no automated bulk migration, on purpose: things move in as they come up, which keeps the filing honest and the first week light.

The core: effectively none. Capture takes zero time (that’s the trick of it), the sweep runs when you say the word, and the weekly system check is about 15 minutes. If you switch on the rituals: five minutes in the morning, five in the evening, 30 to 45 minutes once a week for the personal review. The core works the same at any depth you pick.

Nothing dramatic. Skipped days simply stay empty; nothing nags you, and no streak breaks, because there is no streak. When you come back, you open the folder, say “continue” or “run the steward”, and the next fifteen-minute check picks up the drift: the inbox gets swept, stale items get flagged, and the system is current again. A house waits better than an app: files don’t expire, and the AI reads the state fresh every time.

Both, and the house doesn’t care which. The rooms are the same; what changes is what you put in them: client projects and decisions for a freelancer, the lease and the insurance letters for a household, a side project next to a day job. The three use cases above are drawn on one and the same plan. The personal rituals (morning routine, evening wins, the weekly review) are switches, off by default, and the setup interview offers the ones that fit your week.

Four protections: it must show you a plan before batch edits, in the important folders it may only propose changes, your synced folder’s version history can roll anything back, and every file the sweep moves leaves its untouched original in Archive/Inbox_Originals/. In practice the rules make it boringly careful.

Then you switch, and the house doesn’t notice. This isn’t hypothetical: in June 2026, a US export-control order took two of the newest frontier models offline for every user in the world, Europe included, for almost three weeks. If your context lives inside one provider’s app on a day like that, you wait. If it lives in a folder of plain files, you install another assistant, point it at the folder, and keep working with your whole history intact. And the endgame the house is built for: when local models are ready, the same folder works with them on your own computer, off the big providers entirely.

Personal conviction, honestly held: I believe in one-time purchases. I went through the work of building and running this system, and that finished work is what you’re buying. It doesn’t need a server of mine, an account with us, or a monthly fee to stay alive. And it isn’t the kind of thing that goes stale: the house lives with your development, fills with your projects, your decisions and your people, and gets better the longer you’re in it. I personally believe this is all you need right now.