Privacy Policy

What this policy covers

This policy explains how Norget handles information when you use our website, browser extension, iOS app, imports, Telegram bot, semantic search, memory tree, and AI chat. It describes the information these features process and the providers they may use.

Who is responsible for your data

The data controller is ATAYIL YÖNETİM DANIŞMANLIK HİZMETLERİ LİMİTED ŞİRKETİ, a limited company registered in Türkiye and the company behind Norget. Requests about your personal data can be sent to support@norget.io.

Trade registry: İstanbul Chamber of Commerce, no. 1023586 · MERSIS 0099135693600001

Information we collect

• Account information: your email address, account identifier, the name and profile picture a sign-in provider shares, authentication sessions, preferences, and subscription status.
• Saved bookmarks: the URL, title, page metadata, and page text extracted for reading and AI processing, together with the notes, categories, and collections you create. This includes uploaded files, screenshots, extracted image text and descriptions, AI summaries, search embeddings (numerical representations of text), chat questions, and conversation history.
• Imports and messages: the posts and export files you import, their author and source metadata, and messages and files you send to our Telegram bot. We store raw Telegram updates, including sender and chat identifiers and message content, before processing them, including messages from senders who have not yet linked a Norget account.
• Technical information: request and error logs, IP addresses and browser information handled by hosting services, timestamps, and push subscription identifiers with your browser type, needed to run and troubleshoot the service. Payment records include customer and subscription identifiers and status. For the anonymous public archive question limit, Norget creates a daily keyed HMAC from the visitor's IP address. The raw IP address is not stored in this counter. The counter applies across all public archives, and admitted attempts count even if the answer later fails. A separate per-minute request guard applies to all public archive askers, including signed-in visitors, using a daily keyed HMAC of the IP address without storing the raw address. Signed-in question limits use your account identifier in separate counters across public archives.
• Product events: we record product events together with your plan. When you are signed in, most events are linked to your account. Events about visiting, asking, or reporting someone's public archive are linked to no one, even when you are signed in, and neither are events from visitors who are not signed in. These events cover:
  • When a source becomes available, recorded as source_connected.
  • When at least 20 saved items are indexed, recorded as items_indexed.
  • When you ask a question, recorded as question_asked.
  • When an answer returns: how long it took, which model answered, the answer type, and how many sources it cited, recorded as answer_returned.
  • When you open a cited source, including its rank, platform category, and age range, recorded as citation_opened.
  • When you rate an answer up or down, recorded as answer_feedback.
  • When your home shows saves under “Worth another look”, with how many came from each of its sources and whether AI is on, once a day, recorded as resurface_shown. When you open one, with its source and position, recorded as resurface_opened. Neither records which save it was.
  • When you turn on your public archive, with its public handle, recorded as public_archive_enabled. When you publish items to it, with how many, recorded as public_archive_item_published.
  • When a visitor opens a public archive, clicks one of its buttons, or opens one of its items, recorded as public_archive_viewed, public_archive_cta_clicked, and citation_opened.
  • When a visitor asks a public archive a question, recorded as public_archive_question_asked with the archive's handle and item count, the answer type, how many sources it cited, which model answered, and what the answer cost. When a visitor reaches a question limit across public archives or a limit for an archive, or a question guard is unavailable or the owner has disabled AI, recorded as public_archive_cap_hit with the handle, item count, and reason. When a visitor reports a public archive, recorded as public_archive_reported with the handle and the reason chosen.
  • When someone runs the dead-link check, with how many links it checked and how many were dead, recorded as graveyard_check_started and card_generated. When someone views, shares, or follows a link on the result card, with counts and the share channel, recorded as card_landing_viewed, card_shared, and card_cta_clicked.
  • When you import a file, from the dashboard or from the dead-link check, recorded as import_started and import_completed. These carry the file's source type, how many items it held, how many were imported or skipped, whether a Pocket bonus applied, and whether your plan's limit stopped the import.
  • When someone opens the Pocket page or clicks one of its buttons, recorded as memorial_viewed and memorial_cta_clicked.

  We use these events to measure whether answers are useful and whether the public pages work. They do not include your question text or bookmark content.

How we use your information

We use your information to authenticate your account; save, sync, organize, and search your bookmarks; create summaries, categories, and embeddings; answer questions in AI chat; show connections in the memory tree; maintain security; and diagnose service problems.

Where GDPR applies, the legal bases are performance of our contract for requested features, legal obligations for required records, and legitimate interests in security and service reliability, subject to your rights. Where KVKK applies, we rely on the corresponding conditions in Article 5, including necessity for a contract, legal obligations, and legitimate interests that do not harm fundamental rights. Where consent is required, we must obtain it separately; accepting this policy is not blanket consent.

AI processing and service providers

Chat, answers, summaries and categorization first use a model Norget operates itself. We run the open-weight 0GM-1.0 model, a Qwen fine-tune, on hardware owned by ATAYIL YÖNETİM DANIŞMANLIK HİZMETLERİ LİMİTED ŞİRKETİ. Requests can include your question and relevant saved text, titles, URLs, descriptions, notes, categories, author/creator handles, sender names, save dates and conversation context.

Norget reaches this self-hosted machine through an encrypted Cloudflare Tunnel protected by Cloudflare Access. Cloudflare processes traffic and related service data to provide the connection and access controls.

If the self-hosted model is unavailable, times out, returns an empty or invalid answer, or fails with certain errors such as rate limits or server errors, Norget may send the request to OpenRouter, our hosted processor, for fallback and ask for DeepSeek V4 Flash. Other errors, such as a rejected request, stop without a fallback. OpenRouter picks the company that hosts the model for each request. That can be DeepSeek itself, whose servers are in the People's Republic of China, or another company that hosts the same model. OpenRouter and that company process the request under their own terms and privacy policies.

Answers on a public archive page work differently. When someone asks a question on a public archive page and Norget finds matching published saves, it sends the question, the archive's name, and the text, titles and public notes of those saves straight to OpenRouter for DeepSeek V4 Flash, without trying the self-hosted model first. If nothing matches, Norget says it found nothing, without calling OpenRouter.

Search embeddings are numerical representations of text. To find matching saves, Norget turns saved text, search queries and questions, including questions asked on a public archive page, into search embeddings on its own self-hosted machine with Qwen3-Embedding.On 28 October 2026, we switched to this model and replaced the earlier OpenAI embeddings by re-embedding the saved text. Since that cutover, Norget does not send saved text or queries to OpenAI for embeddings.

Image and screenshot reading goes straight to OpenRouter, which uses Google Gemini 2.5 Flash-Lite. OpenRouter and Google receive the image and instructions for reading it.

Norget does not use your content to train models.

• 0G Compute: an optional second provider after the self-hosted model and before the hosted fallback. If configured, Norget would send 0G Compute your question and relevant saved text, titles, URLs, descriptions, notes, categories, author/creator handles, sender names, save dates and conversation context. Norget would request private inference in a trusted execution environment (TEE). Norget does not independently verify the TEE. 0G Compute is not enabled today, and no content is sent to 0G.
• OpenRouter: our hosted processor for chat fallback, public archive answers and image reading. For each request it picks the company that hosts the requested model. OpenRouter and that company process the request under their own terms and privacy policies. See OpenRouter's terms and privacy policy.
• DeepSeek, or another host of DeepSeek V4 Flash: DeepSeek V4 Flash is the text model Norget requests through OpenRouter for chat fallback and public archive answers. OpenRouter may send the request to DeepSeek itself, in China, or to another company that hosts the same model. These requests can include the question and relevant saved text, titles, URLs, descriptions, notes, categories, author/creator handles, sender names, save dates and conversation context. See DeepSeek's terms and privacy policy.
• Google/Gemini: Google hosts Gemini 2.5 Flash-Lite, the model Norget requests through OpenRouter for image and screenshot reading. It receives image bytes and reading instructions. Extracted text can then reach the self-hosted chat and embedding models, or the hosted chat fallback.
• Cloudflare: Tunnel and Access carry and protect requests to the self-hosted chat and embedding services. Cloudflare processes traffic and related service data under its data processing addendum.
• OpenAI: historical embedding provider only. OpenAI processed saved text and queries for embeddings before the 28 October 2026 cutover. We have replaced those embeddings with Qwen3-Embedding vectors.
• Supabase: authentication, database, and private file storage handle account data, saved content and uploads, embeddings, AI results, conversations, product events, raw inbound messages, integration records, and notification preferences.
• Optional sign-in with Google, Apple, or X (Twitter): when enabled and chosen by you, the provider handles your authentication request and shares an account identifier, authentication tokens, and the email or profile details it makes available with Supabase and Norget. Native Sign in with Apple passes an Apple identity token to Supabase. When Apple token revocation is enabled, the iOS app also sends Apple's one-time authorization code to Norget, which exchanges it with Apple for a refresh token, stores it encrypted, and uses it only to revoke Norget's access when you delete your account. X sign-in requests bookmark-read and offline-access permissions as well as account and post-read permissions; it is separate from the currently disabled bookmark-sync connection described below. Providers process your sign-in under their own policies.
• Vercel: hosting and server functions process web requests and the content submitted through them, along with network information and operational or error logs. We have no separate Sentry, PostHog, or Google Analytics integration in the application.
• Payment processor: our merchant of record receives your email, Norget account identifier, selected product, and the billing and payment information you provide. Norget receives subscription and transaction status. The processor handles card details and records for its own payment, fraud-prevention, and tax obligations.
• Telegram: if you use the bot, Telegram handles messages, sender/chat identifiers, attachments, and the replies we send, including AI answers and saved-item details. Bot questions use the same chat and embedding services as the web app; images can use the configured vision service. If you ask for first-week reminders (with /remind or in Settings), we also send that chat, on days 1, 3 and 7 after a save during your first ten days, the titles of up to three of your own saves with a link into Norget; never the note, summary or page text. /stop or the Settings switch ends them on every channel. Telegram retains its own copies under its policies.
• Resend: when email delivery is configured, it receives your recipient address and notification or digest content, which can include bookmark titles, links, summaries, reminders, and unsubscribe links. Authentication emails are handled through Supabase's configured email delivery service.
• Plausible: when analytics are enabled, the cookieless analytics service receives page-visit information such as page address, referrer, browser, and network information. It sets no cookies. Page addresses can include sign-in, registration, or dashboard landing pages; the configured exclusions stop page-visit reports from dashboard subpages and authentication callback paths. The exclusions do not cover events. On the dashboard, the first-run guide sends Plausible onboarding_shown, import_source_chosen with the import source you picked, first_item_ready, first_question_asked, and first_answer_cited. Each event carries its name, the page address, and the browser and network information a page visit sends, but no account ID. It is not used to send saved bookmark content or chat messages.
• Google Fonts and notification services: browsers request fonts from Google, exposing network and browser information. If you enable web push, your browser's push service receives a subscription endpoint and encrypted notification payload for delivery. Norget stores the subscription and its encryption keys. If you turn on first-week reminders in Settings for push, a reminder is one such notification: the title of one of your saves and a link into Norget. Turning push on alone does not send them. We record each reminder we send (which saves, which day, which channel, and whether it was delivered) so none is sent twice.

Importing or forwarding content can trigger the same AI processing as saving it directly. When we retrieve a saved URL or your browser loads source images, the source website receives the request. Service providers may use their own subprocessors. Their retention and model-training practices depend on the service, routing settings, and applicable contract; we do not promise zero retention or a universal training opt-out.

International transfers

We are based in Türkiye. The services above may process data outside Türkiye and the European Economic Area, including in the United States. Hosted chat fallback, public archive answers and image reading go through OpenRouter, and OpenRouter picks the host for each request. A text request can be processed by DeepSeek in the People's Republic of China, which describes processing and storage in China in its privacy policy, or by another host in the United States or another country. We cannot tell you in advance which country a given request will be processed in. For the other services, locations depend on the service and its hosting region.

For transfers subject to GDPR, a valid basis is required: an applicable adequacy decision, or appropriate safeguards such as the European Commission's Standard Contractual Clauses (SCCs), with a transfer assessment and supplementary measures where needed. For transfers subject to KVKK Article 9, an applicable adequacy decision or an appropriate safeguard is required for routine transfers. If a KVKK standard contract is used, it must be notified to the Turkish authority within five business days of signature. EU SCCs alone do not establish compliance with KVKK's separate requirements.

Ask support@norget.io for the recipients, countries, transfer basis, and a copy or description of the safeguards that apply to your data. See the European Commission's transfer guidance and KVKK's transfer guidance.

Security and encryption

Access to private records and uploads is controlled through Supabase authentication, database access policies, and a private upload bucket. The X connection implementation encrypts stored tokens using AES-256-GCM with per-user derived keys. AI features require content to be readable by the configured processing service; your archive is not end-to-end encrypted against Norget and those services. No internet service can guarantee absolute security. Essential authentication cookies and local session storage keep you signed in; they are separate from optional analytics.

Browser extension and social imports

When you save a page with the Norget browser extension, it sends the tab's address and title to your Norget account. It sends the readable text of the page only if you tick "Archive page content"; otherwise Norget's server fetches the page itself to extract its text. The text is stored, because storing it is what lets Norget answer questions about what you have read and quote the passage it answered from. Long pages are kept up to a length limit rather than in full.

Instagram import is session-based: after you start a save or bulk import, the extension reads posts and fetches captions using your existing signed-in Instagram browser session. Instagram import is not OAuth, and the extension does not send your Instagram password or session cookies to Norget. It sends the selected post URLs, captions, and author metadata. When a bulk import finishes, it also sends your Instagram username, taken from the page address, and the number of posts it imported. Norget keeps both with your connected accounts.

X (Twitter) bookmark import uses your signed-in browser session. Before it begins, the extension asks for your explicit consent and waits for you to choose "Agree and start import." It then scrolls your signed-in X bookmarks page automatically, up to 300 scrolls in one run, and reads each bookmark X loads. It sends those posts' URLs, text, author information, and available timestamps to Norget. It does not send your X password or session cookies to Norget.

A visible "Stop import" control stays on the X page if you close the popup. When you stop, the extension stops reading, scrolling, and starting new uploads. A batch already sent may finish, and items already saved remain in your account. This is not background sync and it may not find every bookmark. The separate X OAuth 2.0 PKCE bookmark-sync connection is currently disabled for new connections; it is not required for the browser import. That separate connection, if enabled, obtains access and refresh tokens for X's bookmark API. File import also accepts Instagram exports and supported X/Twitter archive files such as likes; an X archive is not a complete export of your X bookmarks.

Chrome bookmarks import runs only after you click Chrome in the popup and allow temporary bookmarks access. The extension reads each HTTP or HTTPS bookmark's URL, title (up to 500 characters, when present), and original save date (when valid) and sends those fields to /api/bookmarks/batch with source chrome_bookmarks. It removes bookmarks access after the import finishes or fails and does not track later bookmark changes. A pending intent id and expiry time valid for two minutes, an interruption cleanup marker, and a retry cursor containing only a SHA-256 fingerprint and numeric offset stay in browser session storage; no raw bookmark data or sign-in token is stored there. Local extension storage keeps only the last run time and imported count for this import.

The extension adds controls on X and Instagram and processes content when you choose to save or import it. When you open the extension while signed in, it sends the address of the tab you opened it on to your Norget account, only to check whether you have already saved that page. That check does not add the page to your archive. Beyond that, it does not upload your browsing history. Social platforms continue to process your activity under their own policies.

Writing suggestions run only when you use the Find my saves hotkey or menu command. They require explicit consent. If consent is off or the host is muted, nothing leaves your browser. When you invoke the feature, the extension sends the selection, or text from the focused field, up to 2,000 characters, together with the hostname.

Writing suggestions send draft context to an embedding endpoint only when the resolved host is on an operator-managed allowlist. The operator lists only Norget's self-hosted Qwen3-Embedding host. If the allowlist is missing, the resolved host is not listed, or the formatted input exceeds the embedding budget, Norget uses keyword-only retrieval. Draft context never goes to OpenAI, a hosted embedding fallback, or a completion model. We do not store the draft text, and there is no option to opt in to raw draft storage. We store a per-user HMAC of the normalized draft context, character count, source, hostname, client version, and request timestamp. We keep a monthly suggestion usage count until account deletion. The count includes requests that Norget later refuses because the global spending limit was reached. We also store whether a suggestion was shown, accepted through copy, insert, or open, or dismissed with an optional reason: irrelevant or wrong project. These records link the request ID to the bookmark ID, rank, and timestamp.

Retention and deletion

Your archive, embeddings, generated metadata, uploads, and conversations are stored to provide the service while you keep them in your account. Cancelling a paid plan does not delete your archive. You can delete bookmarks and conversations, or delete your account from settings. Account deletion removes your stored uploads before deleting the account and database records linked to it. If you signed in with Apple and Norget holds an Apple token for you, Norget asks Apple to revoke it as part of account deletion; otherwise settings shows you how to remove Norget from your Apple Account yourself. Product events stay after deletion, no longer linked to any account. Events about a public archive keep its public handle.

Deleting your account does not cancel a paid subscription. If you have a subscription, email support@norget.io and we will cancel it for you.

Raw Telegram messages have no automatic expiry in the application. Deleting a bookmark removes its uploaded file but not the original Telegram message it came from. Messages from unlinked senders may not be associated with an account, so contact support to request their removal. Disconnecting a channel stops its use for your account; it does not erase previously saved content or copies held by Telegram.

Writing-suggestion request metadata remains until you delete your account. Deleting a bookmark removes its suggestion events and provenance links, but not the request metadata.

With each anonymous public archive question, Norget removes up to 1,000 question-limit counter rows dated earlier than two days before the current UTC date. A backlog, temporarily locked rows, or little or no new traffic means old counter rows can remain longer.

For request-guard and signed-in question counters, windows older than two days are eligible for cleanup on a one percent random sample of rate-limit requests. This cleanup is not scheduled, so those counters can also remain longer when there is little or no traffic.

Backup copies, operational logs, and provider-held records follow the relevant service's retention arrangements. Billing records may need to be kept for tax, legal obligations, or disputes. Retention periods for these records vary by service, and deletion from backups and third-party systems is not immediate. Contact us for the period or criteria that apply and to request deletion of residual records.

No sale of personal data or advertising trackers

Norget does not sell personal data. We do not use advertising trackers or build advertising profiles from your bookmarks, searches, connected accounts, or extension activity.

Your choices and rights

Settings provides a bookmark JSON export, notification controls, and account deletion. For a broader export, access, correction, erasure, restriction, portability, objection, or withdrawal of consent where applicable, email support@norget.io. We may need to verify your identity. Withdrawal does not affect earlier lawful processing. Under KVKK you may also ask whether data is processed, its purposes and recipients, request notification of corrections or deletion to recipients, object to adverse results produced solely by automated analysis, and seek compensation for unlawful processing. Rights are subject to the applicable legal conditions.

We respond without undue delay: within 30 days for KVKK requests and normally within one month for GDPR requests, with notice of any legally permitted extension. You may complain to the Turkish Personal Data Protection Authority under KVKK's complaint procedure or to your competent data protection authority under GDPR.

Changes and contact

We may update this policy as Norget changes. We will update the date at the top when we do. Questions about this policy or your data can be sent to support@norget.io.