What leaves your Mac: the data path of private AI, memory and meeting apps
Apps differ on five layers: capture, transcription, search, AI and sync. Everpaper, Obsidian, Anytype and Screenpipe keep the core record on your device by default. Granola, Otter, Fireflies, Mem and Notion process in a cloud. Day One keeps entries end-to-end encrypted on every plan. What Everpaper itself sends is limited to model downloads, update checks and the AI providers, embeddings providers and agents you connect.
Disclosure. Everpaper is our product, and it is in this article. We say where it falls short, and every competitor fact below carries the date we checked it.
Facts as of October 2026. Prices and features change. Each source is linked at the end.
“Private” gets used for any app with a lock icon. A more useful test is what happens to your data in the default setup, layer by layer: capture, transcription, search, AI and sync. An app can be local for four layers and cloud for the fifth.
Default data path
| App | What it says about the default path |
|---|---|
| Everpaper | Capture, transcription, search and the journal on your Mac. AI only through a provider you set up; cloud providers need a switch that is off by default |
| Screenpipe | Capture history stays on your device by default; cloud AI, sync and integrations use separate paths. Its README says product analytics (PostHog) is on by default and Sentry receives crash diagnostics while telemetry is enabled |
| Obsidian | Notes are local files and the app needs no sign-up; Sync is optional at $4 a month billed annually |
| Anytype | Local-first, peer-to-peer and end-to-end encrypted, per its README |
| Day One | End-to-end encryption on all plans, including free |
| Granola | Audio not stored; transcription and notes through providers such as Deepgram, OpenAI and Anthropic; trains on anonymized data for non-Enterprise users unless they opt out |
| Otter | Cloud, with Amazon Web Services in the United States; policy lists training its AI on de-identified audio and transcriptions |
| Fireflies | Cloud; policy says it does not use personal information for AI model training |
| Mem | Cloud; encrypted at rest and in transit, not end-to-end encrypted |
| Notion | Cloud workspace; offline mode on every plan; zero data retention with AI providers is an Enterprise feature |
| Windows Recall | Off until you opt in; snapshots encrypted on the local drive; not shared with Microsoft per its documentation. Windows only |
Local and end-to-end encrypted are different
Local means the vendor never receives the data. End-to-end encrypted means the vendor stores data it cannot read. Both beat a plain cloud. A local app depends on your Mac and your backups. An end-to-end encrypted service depends on its client code and on you keeping the key.
What Everpaper sends, in full
Everpaper is our product. Its privacy page lists what connects to the internet:
- Model downloads when you set up speech or AI models, from Hugging Face and everpaper.app, and the oMLX server from GitHub.
- Update checks to everpaper.app, daily by default and never during a meeting recording. You can change this or turn them off.
- Your AI provider, and an external embeddings provider if you set one.
It has no analytics or telemetry. Recordings, screenshots and your calendar are never uploaded, and only text such as screen text, transcripts and meeting titles can reach an AI you set up.
AI is opt-in. With no provider, AI features wait and nothing is sent. On-device providers (Everpaper Native on macOS 27, oMLX, LM Studio, Ollama while on this Mac) keep text on the Mac. Cloud providers (Anthropic, OpenAI, Google Gemini, OpenRouter) get text only after you turn on Allow cloud AI to analyze your journal, which is off by default. A private endpoint you mark as trusted receives text without that switch, and two other optional connections exist: an external embeddings provider and Allow external agents (MCP). Keys, card numbers, Social Security numbers and similar secrets are scrubbed on the device before any call, and for cloud providers it also replaces emails, phone numbers and names it detects.
What this does not cover: a screenshot is an image, so a secret visible on screen is still in it. Screenshots stay on the Mac and are deleted after 30 days by default. Computer History, which records clicks and the text you select or type, is off until you turn it on, and typed text is kept as an encrypted, redacted record whose original is gone after 48 hours.
How to choose
- The record must not leave the device: Everpaper, Screenpipe, or Obsidian for notes you write.
- You want sync and the vendor must not read it: Anytype or Day One.
- You accept a cloud for the features: read the policy line above for the vendor you pick.
For meetings specifically, see who hears your meetings. The full Everpaper details are on the privacy page.
Apply for the Everpaper private beta.
Frequently asked questions
What is the most private AI memory app for Mac?
Judged by default data path, the ones that keep capture, transcription and search on the device: Everpaper, and Screenpipe, whose README says capture history stays on your device by default. Obsidian keeps notes as local files but has no capture of its own.
Does Everpaper send anything to the cloud?
Only model downloads, update checks, and the AI provider you set up. It has no analytics or telemetry. Recordings, screenshots and your calendar are never uploaded. Cloud AI providers receive text only after you turn on a switch that is off by default.
Do any of these apps train models on my data?
Granola says it trains on anonymized data for non-Enterprise users, who can opt out in Settings. Otter's policy lists training its AI on de-identified audio and transcriptions. Fireflies' policy says it does not use personal information for AI model training.
Is end-to-end encrypted the same as local?
No. End-to-end encryption means the vendor stores data it cannot read. Local means the vendor never receives it. Anytype and Day One advertise end-to-end encryption, and Mem says it is not end-to-end encrypted.
Sources
Written by Marco Iannello Updated 21 October 2026 How we write comparisons