Building your vault

Start in 30 seconds

No export yet? On the Start page, Tell Context OS about you in 30 seconds: paste your ChatGPT memory or custom instructions, write a short bio, or dictate it. Context OS distills it into a live SOUL.md and AGENTS.md right away. Already have an export? Drop it into the free audit instead; Export your AI history shows how to request one.

Your text stays in the box until the build has distilled it, also if a build fails or you reload the page, and the build shows each step as it goes. When it is done you land on Here's what Context OS learned. Your vault is stored in compact English for your agents; Vault Chat answers in your language.

Want to look around first? The demo opens on a sample vault you can browse (its distilled files, SOUL.md and stats); click Build this as my vault (free): the demo becomes a sample vault instantly, at no token cost, and it doesn't count toward your plan's vault limit. The sample is read-only demo data (Layla Haddad's): ask it a few questions a day without using your tokens, but connect your agents and add your own context in your own vault.

After a build you see what Context OS learned: facts, projects and workflows, your SOUL.md, three questions to ask your vault, and one click to connect Claude Code, Cursor or VS Code. The screen celebrates the moment your agent first reads the vault.

Supported sources

Sources can be in any language, and mixed-language uploads are fine. Vaults store compact English by default; Vault Chat answers in the language you write in.

The free audit and the Distill Context dialog accept .zip, .json, .html, .pdf, .docx, .csv, .tsv, .md, .markdown and .txt files. A file of another type is named right away, with the list of accepted types, instead of disappearing. CSV and TSV tables are read in parts of 500 rows with the header repeated, so no row is dropped. A scanned PDF has no text to read: run OCR on it or export it with its text, then add it again. If the download of an export ZIP was interrupted, the files that arrived are still read, and the ones that could not be read are named.

What happens during a build

1. Delta sync. Sources you already distilled are skipped, so you never pay twice for the same conversation. When a conversation continued since your last export, only its new messages are distilled.
2. Free pre-cleaning. Trivial chats, system noise and very long code blocks are stripped before any AI call. This usually removes 70–85% of the raw tokens.
3. Secret scrubbing. API keys and tokens, passwords (also quoted ones and ones labelled in Arabic), webhook links, credentials in URLs and commands, private keys, wallet seed phrases, card numbers and IBANs are replaced with placeholders in your browser, in conversation titles too. The real values stay on your device.
4. Chronological distillation. Conversations are processed from oldest to newest, and each fact is dated by the message that said it: something you said in 2025 inside a conversation started in 2023 still beats what you said in 2024. When a newer fact contradicts an older one, the newest truth stays in the file and the older one moves to a dated history section. Every fact remembers which conversation or document it came from, so agents can cite it.
5. Tidy folders. Procedures and notes grow with each build instead of being overwritten, and once a build is done, the description and routing rule of each folder it changed are refreshed from what the folder now holds.

Vault Chat finds the right files across languages: a question in Arabic finds notes stored in English, and the other way round. Stop answering ends an answer at any time, and answers follow the style your vault records (for example short answers with one recommendation).

While a build runs, the vault shows its progress batch by batch ("Batch 3 of 12"); you can leave the page and get a notification (in the app and by email) when it is ready, with how many facts were learned. If something goes wrong, the reason is explained in your language, and a damaged or unsupported file is named instead of being skipped silently.

If your token allowance runs out mid-build, the newest sources that fit are distilled first. The older ones stay staged, encrypted, for 7 days (up to what one Import Pack can distill): buy the Import Pack or upgrade to Pro, then click Distill my staged conversations now. Nothing is distilled without that click, so a purchase never spends tokens on its own. You never upload the export a second time. A new build needs tokens left (or your own AI key): with an empty allowance, you are offered more tokens before anything is uploaded. Nothing is marked as processed until it actually is. Large imports are saved in steps, so a problem late in a build never undoes what was already distilled, and very large imports continue on their own until they are done. If the allowance runs out in the middle of a step, everything distilled up to that point is kept. If the AI returns an answer it cannot read for part of an import, it tries once more and otherwise skips that part (not marked as processed, so uploading it again retries it), and the rest of the build goes on. Very long documents are uploaded in parts instead of being cut off.

Voice notes can be up to about 6 minutes (3 MB) long; recording stops at that limit. If a transcription fails, you can retry it without recording again.

Folders and files

Every vault starts with the same numbered hierarchy, so sorting is predictable:

• 01_Personal_and_Professional_Context
• 02_Business_and_Projects
• 03_Code_Projects
• 04_SOPs_and_Skills
• 05_Integrations_and_Tools
• 99_Temp_Workspace (drafts older than 7 days are cleaned up)

New topics get their own folders (06_ and up), and so does a folder you create at the top level: it gets the next free number (for example 06_Pricing), so it sorts with the other topic folders and before 99_Temp_Workspace. Type a two-digit prefix yourself to choose the number. Each file carries YAML frontmatter with a stable ID, tags and a three-line TL;DR. Rename or move files freely: the routing maps, and the links in the history archive, update themselves at no token cost.

In the Explorer, the generated maps sit together under Maps for your agents, and empty folders are dimmed. The editor keeps unsaved changes when you switch files or tabs, marks such files in the tree and asks before you close the page. Images, HTML and comments are kept exactly: in Rich text they show as read-only Markdown blocks you edit in the Markdown tab.

Routing maps

Context OS generates AGENTS.md, CLAUDE.md, GEMINI.md, SOUL.md, .cursorrules, .windsurfrules and a README.md per folder. They contain the repository map, IF/THEN routing rules ("if the task is about pricing, read 02_Business_and_Projects") and self-maintenance instructions, so agents read only the folder that matches the task.

Quarantine

Files marked sovereign (your core identity and strategy files, for example) are never overwritten silently. Agent edits to them, and facts that could not be classified with confidence, wait in Quarantine as a diff (side by side on large screens, one column on phones) that names the token or share link behind each proposal. Approve or reject each change with one click.

Staying fresh

• Staleness Radar flags core files you have not confirmed in 60 days.
• Contextual Magnet: create a new folder and Context OS proposes the files and sections that belong in it, with one-click undo. It only moves files and sections that everyone who can open the new folder may already read, and the result says how many items stayed where they were.
• Voice notes route a quick spoken update to the right file and log it in the changelog.