Free · Open source · The memory the AI reads by itself

What if your AI already knew
where you left off?

A free folder you paste into your assistant. From then on it knows who you are, what you are working on and what you decided yesterday, without you explaining any of it again. Open a new chat and it picks up where you left off.

Free. No email, no account. To uninstall it you delete a folder.

Forpeople running several projects at onceWorks withClaude, ChatGPTVersion1.4, in English

    AI today

    no context

     

    With Context_OS

    same model

     

    Example projects, answers written out rather than recorded. The sources it cites are text files in your own folder: no software, no account. Tap a question to watch it.

    You already have an AI that reasons well. You are still the one supplying the context.

    Open a chat. It doesn't know who you are, what you're working on, what you decided yesterday. You explain everything again. Tomorrow you start over.

    The more projects you run in parallel, the more this costs: time, tokens, and decisions that get lost because they stayed inside a closed conversation.

    The proof is one screenshot.

    A new chat, days later, in a conversation that knows nothing about the ones before it. I told it nothing: it read.

    A new session listing active projects and open threads on its own, without anyone telling it

    It installs with one prompt. Literally.

    1.

    Create an empty folder, wherever you want on your computer.

    2.

    Point your assistant at it: in Claude Code or Cowork, select it as the working folder.

    3.

    Paste this and send it:

    The prompt to paste
    Download the zip of https://github.com/Lucas-Grilli/Context_OS (branch main) and extract its contents directly into this folder. Then delete the downloaded zip. Finally open CLAUDE.md, read the section "First action of every session" and follow it.
    4.

    You are up and running. The system notices on its own that it is empty, asks you 3+1 questions, and works with you from there. Three minutes, timed.

    The install prompt pasted into the assistant, a moment before sending it

    That's it. Do you work in Italian? The same folder exists in Italian.

    Why it holds up as the folder grows.

    The hard part of an archive isn't reading: it's not re-reading everything every time. The rules file never holds state, only the map to reach it. A request goes from the map to one index to one file, and the rest of the folder stays closed, even at hundreds of files.

    A request travels one single path: from the rules file to an index to a file, the rest is never read

    Rules, index, file: one path read per request.

    Seven procedures it runs, not ones you read.

    7
    procedures ready
    3
    minutes to start
    0
    accounts to open
    2
    languages available
    Onboardingthree questions on first run, and the system configures itself
    Librariansorts raw notes between knowledge and projects
    Project kick-offfolder, rules, state and index: nothing is born orphaned
    Knowledge pilla skill explained once, callable forever
    Specializefetches the evidence before answering off the cuff
    System checkbroken links, orphan files, rules drifting from practice
    Session logcloses the day with two dated lines in the log

    The three things you are thinking.

    “I already have Notion, or Obsidian.”

    Those are for you to re-read, which is exactly why you stop after three weeks. Here the reader is the assistant: it opens the folder by itself, before it answers you.

    “The AI already has memory.”

    It decides what to keep and you never see it, and it stays inside the product. Here the files are yours, you open them in any editor, and they stay yours even if you switch assistant.

    “It will turn into a second job.”

    The boring part — keeping indexes, knowing what sits where — is exactly the part it does. If you end up playing librarian by hand, the system has failed.

    Who it is for

    →You run several projects at once, and every chat starts with you explaining who you are.
    →The decisions you make with the AI get lost inside conversations you have already closed.
    →You tried a second brain and abandoned it because maintaining it cost more than it returned.

    Not for someone with a single project they keep entirely in their head: that context is already there.

    What it doesn't do

    ✕It doesn't sync across devices: it's a folder, use whatever sync tool you prefer.
    ✕It doesn't import the notes you already have elsewhere: you bring them in when they matter.
    ✕It isn't built for several people on the same system: it's personal.
    ✕It was built and tested on Claude: elsewhere it's plain text, but the flight hours are there.

    It isn't a demo. It's the folder I work in.

    What you download is the skeleton of the system I run my own work on every day: same rules, same procedures, without my content inside.

    My archive today holds 553 files and over 465,000 written words. A normal session reads a tiny fraction of it, because there is a map telling it where to look before it starts looking. That is the only difference that matters between an archive that speeds work up and one that slows it down.

    Every correction I make to it goes back into the system, and from there into the version you download. That's why it is at version 1.4 and not 1.0.

    Try it, and tell me where it breaks.

    Worst case you lost three minutes, not money.

    Code on GitHub