Privacy
Last updated 15 November 2026
Who runs Trove
Trove is run by Michael Litman in London. For privacy or security, email hello@mikelitman.me. There is no separate security address, no phone number and no postal address on this site.
What is stored
- Your account: email address, and a password stored scrambled by Supabase Auth.
- Each save: the link, the page text Trove fetched, your notes, why you saved it, highlights and the notes on them, tags, collections, and fields the model wrote such as a summary, themes and a cultural read.
- A transcript, when one was made, and any text read out of an image. Both sit on the save.
- Narration audio, images you uploaded, and library files (images, PDFs, decks and documents) in private storage. A file is kept even when its text is not read.
- Inbox items from feeds you follow and newsletters you forward, drafts you saved, a writing sample if you added one, your listen queue, and Signal radar themes built from saves that are not kept private.
- Billing status. Card numbers stay with Stripe.
- Which AI apps you connected, the access you chose, and a log of which Trove tools they called. API keys are stored so the app can check them. An export lists the label and scope, not the secret.
Signal radar groups those saves into themes and stores the theme, a short thesis, and which saves belong to it. A nudge on the briefing page can be a notice computed from theme counts, or a stored notice the daily job writes from your own saves. Reading position and type size stay on your account. A daily highlight review is stored on your account and stays in the app. The owner of Trove can receive a weekly note of his own titles, themes and domains on ntfy.sh. Other people's libraries are not posted there. The iPhone Shortcut uses a save-only key: it can add a save and cannot read the library back.
What the AI reads, and when
A save is sent to a model only for a job you turned on by using the product. The usual path is:
- After you save, OpenAI writes a summary, themes, entities, tone and highlights, then an embedding (a vector) so search by meaning can find it.
- If the save is an image, or a page that is mostly an image, one OpenAI call reads the picture. The text is stored on the save.
- A cultural read is a later OpenAI call. It can include public comments Trove fetched from YouTube, TikTok, Reddit or X. You can turn that read off per save.
- A video or podcast uses captions when Trove can fetch them. If it cannot, the audio address is sent to Groq and the transcript is stored.
- An uploaded file is read for its text, and an image is read by the same vision call, unless that save is kept private. The file itself stays in storage either way.
- Signal radar sends titles, domains and theme words from the saves in a cluster to OpenAI when that cluster needs a name. Suggested moodboards do the same with titles and tags of pictures that share a look. Insight briefs send the saves you picked, after private ones are dropped. The weekly what's shifting brief sends this week's saves the same way, and can also send your notes, highlights and writing sample when you ask it to write in your voice.
- A stored nudge can send a note you wrote, and the summaries of later saves, to OpenAI once a day to check whether they contradict. Private saves are left out before that call. The reader itself does not send the article to a model.
- Same Vibe compares pictures. The colour thumbnail is calculated in Trove and is not sent out. When a Voyage key is set on the deployment, the picture is also sent to Voyage and the vector is stored. Look search sends a description you type, or a picture you drop, to Voyage and compares it with those vectors. It does not store the query. Search across formats uses that same look search for a visual query. A save kept private is not sent and is not returned. Moodboards cluster the stored vectors inside Trove and do not send the pictures again.
- Ask, the weekly briefing, the what's shifting brief, taste notes, recommendations, connections and writing each send titles, summaries, themes or excerpts to OpenAI when you open them or when the briefing email is built.
- Listen sends the article text to ElevenLabs only when you press it and there is no stored narration. The file is kept so the next play does not call ElevenLabs again.
- A connected app, through MCP, reads the saves its tools return. A briefing tool also calls OpenAI.
Mark a save Keep private from AI and it stays in your library for you. It is left out of those model calls, embeddings, Voyage, Same Vibe results, look search, moodboards, briefings, insight briefs, the what's shifting brief, Signal radar, Ask, writing, transcription, image reading, file-text extraction, cultural reads, new narration, nudges, the monthly taste note, connected saves, the daily highlight email, Readwise and Notion sync, the public Patina page, and answers a connected app can fetch. Search by keyword inside the app can still find it, and a vault you download still includes it. Settings has the same choice as the default for new saves, including files you upload. Changing the default does not rewrite saves you already have. Turning the flag off queues the usual reading again, including text from a file that was not read while it was private, and a Voyage vector if that key is set.
Who else handles it
The database project is in the EU (Frankfurt). That is Trove's stated Supabase location. The site and its server functions run on Vercel, so text that a function reads passes through Vercel. This page does not state a Vercel region or a log-retention period, because the code does not set one.
- Supabase stores the database, sign-in, and the audio and image files. Supabase privacy.
- Vercel hosts the site. Vercel privacy.
- OpenAI does the chat, vision and embedding calls above. The chat and vision model in this code is gpt-6-luna. Embeddings use text-embedding-3-small. OpenAI's API documentation says data sent through the API is not used to train models unless the organisation opts in. This code does not opt in. OpenAI may keep abuse-monitoring logs for up to 30 days. Your data and how data is used to improve models.
- Groq transcribes audio when captions are missing (whisper-large-v3-turbo, unless the deployment sets another model). Groq's services agreement says it is not permitted to use inputs or outputs for training or fine-tuning unless the customer grants permission. This code does not grant that. Groq says inference is not retained by default, and that reliability and abuse logs can be kept for up to 30 days. Zero data retention exists as a Groq option. This code does not turn it on. Services agreement and Your data.
- ElevenLabs speaks an article when you press Listen (eleven_turbo_v2_5). ElevenLabs says self-serve plans can use data to improve models unless that account turns the setting off. This code does not turn it off, so this page does not say your narration text is kept out of training. ElevenLabs data help.
- Voyage receives a picture for Same Vibe, and a look-search description or picture, only when the deployment has a Voyage key set (model voyage-multimodal-3.5). Voyage's privacy policy says customer content may be used to train and improve its models unless that organisation opts out in the Voyage dashboard. This code does not opt out. A save kept private is not sent. Voyage privacy and the opt-out note.
- Resend sends briefing, welcome, collection-share and, if you turn them on, the daily highlight email and a nudge email. A briefing can include titles and summaries. The highlight email can include quotes. Quotes from a save kept private are left out. A nudge email repeats notices that were already built without private saves. If you forward a newsletter, Trove fetches the message body from Resend and stores it. Resend privacy.
- Stripe takes payment. Trove stores the customer id and subscription status, not the card number. Stripe privacy.
- PostHog receives a page view, including the page address, referrer and browser details, sent to PostHog's US cloud at us.i.posthog.com. The browser is set not to write a PostHog cookie or keep the visit on disk (persistence is memory). Autocapture, session recording and heatmaps are off, and paths under /l/ are skipped. The text of a save is not sent. This code does not set how long PostHog keeps the event on its servers, and it does not identify you by email. PostHog privacy.
- Your browser asks Google for a small icon of a saved site's domain. Clearbit is used for a fixed set of logos on the marketing site, not for your library.
Readwise and Notion receive highlights only when you connect them in Settings. The token is stored encrypted and is not included in an export. A save kept private is not included in that sync. Downloading an Obsidian vault is a file for you, and it includes private saves. Exporting one save to Notion with a token you paste sends that save, including one you marked private, because you chose it. Slack receives nothing from the library. If you connect Slack, Trove reads one channel you pick and creates saves from links it finds. The bot token is stored encrypted and is not included in an export.
How long it is kept
Trove keeps your library until you delete a save or the account. OpenAI and Groq describe abuse and reliability logs of up to 30 days in the documents linked above. Voyage does not state a retention period in the policy linked above, except that content sent after an opt-out is deleted once it has been processed. This code does not opt out. Stripe keeps payment records it needs for tax and accounting after an account is deleted. This page does not invent retention periods for Vercel, Resend, ElevenLabs, PostHog or Supabase beyond what those links say and what this code does.
Export and deletion
Settings has a Privacy section. JSON is the library, notes, highlights, transcripts, drafts, inbox, settings and file names. CSV is the saves list. Files is a zip of uploaded images, library files and narration, up to 80 MB in one download. Embedding vectors and API secrets are left out.
Delete account removes your rows, including saves (and the transcript on them), inbox, drafts, the listen queue, highlight review, Signal radar themes, Ask usage, the Slack connection and share-invite history, then the audio, image and library files, then the login. Stripe keeps the payment records described above. A completed invite stays used: the redemption is kept without your account id, so deleting the account does not free that use.
To ask for a copy, a correction, or to object, email hello@mikelitman.me. Trove will reply within a month. You can complain to the UK Information Commissioner's Office at ico.org.uk.
Cookies
Sign-in uses the cookies Supabase needs to keep you logged in. PostHog is configured not to set a cookie. There is no advertising cookie.
Changes
If this page changes in a way that matters, Trove will email you before it takes effect.