Why encrypted is not the same as private
Every privacy page leads with encryption, and encryption is genuinely necessary. It stops outsiders from reading your data in transit and at rest. It says nothing about the company itself. A vendor can encrypt every byte perfectly and still keep your voice notes forever, mine your transcripts for a behavioral profile, or hand the whole archive to whoever buys the company next year.
The rest of the typical policy is permission language. We may collect. We may share with partners. We may retain data to improve our services. Every may is a door the company is holding open for itself, and improving our services is wide enough to drive a training pipeline through.
What protects you is not the strength of the lock. It is the size of the vault. Data that was never kept cannot leak in a breach, cannot be subpoenaed, cannot be quietly repurposed after an acquisition, and cannot train anyone's next model. Discarding data on arrival is the only privacy promise that does not depend on everybody at the company behaving well forever.
The question that matters: what does the app keep?
There is a name for this discipline. European regulators call it data minimization: collect only what the feature needs, keep it only as long as the feature needs it. You can apply it as a buyer without reading a single law. Ask the vendor to sort everything the app touches into three buckets.
Held: what remains on their servers after processing is done. For a task app, that should be your account identity, your list, and your settings. If the honest answer runs longer than a sentence, the app is a warehouse.
Passes through: raw material that is read to produce something and then discarded. The forwarded email, the voice note, the photo of the whiteboard. The output stays. The source should not.
Never exists: data the design avoids creating at all. Passwords replaced by codes that expire. Card numbers handled entirely by a payment processor. A behavioral profile that was never assembled because behavior was never tracked.
A vendor who can produce this inventory quickly has thought about it. A vendor who cannot has usually been keeping everything, not out of malice but because keeping everything is the default, and defaults win unless someone fights them.
Six questions to ask before you connect anything
Does it keep raw captures, or only what it made from them? An assistant that turns an email into a task needed the email once. Ask directly: after processing, does the source persist anywhere, and for how long?
Where does audio go? Voice is the most intimate thing you produce. Ask whether recordings are ever written to disk, whether transcription happens on your device or on a server, and whether a bot joins your calls. A recording bot in the participant list is an upload with a face.
Do its AI vendors keep or train on your data? Nearly every assistant sends text to a model provider. The answer you want is specific: vendors named, one job each, contractually barred from retaining or training. A vague nod to trusted partners is not that.
Can it act as you? An app that sends messages or rearranges your calendar on its own authority is a far bigger risk than one that only drafts and suggests. Ask what it can do without a human pressing the button.
What does delete actually do? Archived is not deleted. Ask how quickly a removed item disappears from every device, what signing out leaves on the hardware, and whether closing your account takes an email or a phone call.
Can you see the inventory yourself? A page that shows everything held about you, with a working delete next to each piece, is the difference between a claim and a receipt.
How Matters answers all six
Raw captures are read exactly once. Whatever you say, type, copy, share, or photograph is turned into tasks and then discarded. When Gmail connections open, they will be read-only: only the tasks are kept, never the mail. On Android, notifications pass the agents through a permission you grant deliberately and can revoke any time, and codes and secrets never reach them at all.
Audio never exists on a server. Voice notes are transcribed and the sound is never stored. Meetings go further: Matters listens on your Mac, on the device, with no bot joining the call, and the audio never leaves the machine. When the meeting ends, the recording does not exist anywhere. The action items do.
Five vendors, named, one job each. Anthropic's Claude reads captures into tasks. OpenAI is the fallback reader and handles speech to text. SendGrid delivers our emails. Bachs processes payment, so your card number never touches us. Vercel runs the server. None of them may keep or train on what they see.
It writes, and you send. Matters drafts the chase-up and the deferral, then stops. It never messages a human on its own and never speaks as you. The same restraint covers your calendar: the brief may say your morning is clear, but Matters never writes blocks onto it or moves your meetings.
Delete deletes. A removed task disappears from every device within moments. Signing out wipes the device. One email deletes the whole account within 72 hours, and the page for doing it has no dark patterns.
And the inventory fits in a sentence: your email address, one copy of your task list so your devices agree, and your settings. Passwords never exist, because sign-in codes are stored hashed and die in ten minutes. Your card number never exists here. A behavioral profile of you never exists at all.
Do AI assistants train on my data?
Some do, and the policy language is where you find out. Look for an explicit statement that neither the company nor its model vendors retain or train on your content, with the vendors named. If all you find is improving our services, assume the answer is yes. The five helpers Matters uses are barred from keeping or training on anything they read.
Does keeping less make the app less useful?
No, because the useful memory was never the raw material. Matters keeps your finished tasks, and Ask can answer from everything you ever completed, months later, when someone wants to know what was agreed. That is long recall built entirely from data you can see and delete. A good memory does not require hoarding the sources.
How do I test any of this?
Pick the promise that matters most to you and try to catch it failing. Delete a task and watch it vanish from your other devices. Sign out and see what remains. Read the Trust Center, then check each line against how the app behaves. Matters is free for 14 days, no card needed, so the test costs you nothing but one messy sentence. Say it, watch the task appear, and know the sentence is already gone.