Your Mac Builds a Hidden Map of Your Family and Contacts, Researcher Finds

What the Graph Holds

The store lives in ~/Library/IntelligencePlatform: 51 SQLite databases, unencrypted, with the main one in a file called graph.db. Two system processes fill it, knowledgeconstructiond and intelligenceplatformd. On macOS 26.6.2, where Pasquini did his analysis in August, the graph’s vocabulary had 181 types of entities and 278 kinds of links between them.

Most of what goes in is copied from apps you already use. Pasquini lists Contacts, the faces in Photos, Calendar, Family Sharing “including the age of every member”, and the emergency contact in your Health Medical ID. For a family with children, that means the kids’ ages too.

Messages and Mail feed it through Siri Suggestions, and apps add their own data through Apple’s App Intents system. A communication log records who you talked to, when, in which app and in which direction. It holds no message content, Pasquini notes, but it “also pulls from your other devices”.

The rest the Mac works out for itself. A relationship label (“e.g. sister, coworker”) comes partly from what Pasquini calls the kinship classifier, an XGBoost model combined with rules. Places get a purpose, such as home or gym. In a follow-up post he shows that the graph can name your employer from calendar invitations and email domains, even when no employer field is filled in anywhere. Some birthdays appear to be estimated from photos, though he marks that one as unconfirmed.

He is careful about the limits of his own work. Each entry in his tables is marked as implemented, unverified or “could not find mechanism”, and he calls the labels “inference work rather than ground truth”. Some declared types may never be filled, and tables come and go between macOS versions. His viewer was tested on macOS 26 and 27, and a Hacker News commenter found the same folder on macOS 15 Sequoia.

Deleting a Contact Doesn't Delete the Conclusion

Remove someone from your contacts and the live graph stops deriving facts about them. What it already concluded stays.

Next to the live store sits a second table, expired_stable_graph, and a retracted fact moves there with its confidence score and its sources intact. “Retraction in this store is an append, not a delete,” Pasquini writes. He estimates that rows leave the table after about 48 hours, yet he could still read very old records in it on his own Mac. One internal scorecard still rated two people the graph had dropped the day before.

The only fix he offers is deleting those rows by hand. He presents it partly as protection against a partner seeing how close the Mac scored you to someone you had removed.

What Apple Documents

Very little. The knowledgeconstructiond man page, dated June 2022, describes “a daemon that analyzes content on the device to build a general purpose knowledge graph”. That is the whole explanation.

Apple’s Siri privacy page says Siri Suggestions “uses local, on-device processing to learn how you use your devices and apps”. With iCloud on, the Siri Suggestions personalization syncs across your Apple devices with end-to-end encryption. Neither page mentions relationship scoring.

Apple researchers described the techniques behind such a graph in a 2023 paper on the company’s Saga knowledge platform, including support for private on-device knowledge. According to Pasquini, it says nothing about how the system is built into macOS or about kinship inference. He writes that AI models retrieve from the store, with an “allegedly” in brackets.

The databases themselves are no secret. A Mac user asked on Hacker News in September 2023 why nobody talked about them, and security professional Nic Lydon looked at the same store in an earlier write-up. Pasquini’s goes through the whole vocabulary, class by class.

Who Can Read It

The graph is built on the device, and the folder sits behind macOS’s Full Disk Access protection, so ordinary apps can’t open it. Any app you have given Full Disk Access can, and so can malware that gets it.

That permission is less rare than it sounds. Backup apps need it to work, and Apple said in October that it will make it harder to grant, citing AI agents that use it to reach files, mail and messages.

Pasquini’s follow-up asks what an attacker who already runs code on the Mac, with Full Disk Access, would get. “A breach hands you records; this store hands you conclusions,” he writes. Inferred relationships with confidence scores, household groupings, faces linked to contacts and their phone numbers, recurring calendar events with locations, communication metadata with resolved names and travel bookings all sit in one place, already joined up.

What Mac Users Can Do

Apple offers no switch for the graph itself, and Pasquini doesn’t name one. A few settings control who can read it and, possibly, what goes in.

• Check Full Disk Access. System Settings > Privacy & Security > Full Disk Access lists every app that can read the store. Remove the ones you no longer use or don’t recognise.
• Keep FileVault on. Macs with Apple silicon or a T2 chip encrypt their data automatically, and FileVault adds your login password as a requirement (System Settings > Privacy & Security > FileVault). That protects the databases on a lost or stolen Mac. It does nothing against apps running on it.
• Limit what Siri learns from. Under System Settings > Apple Intelligence & Siri > Siri Suggestions & Privacy, you choose the apps Siri Suggestions works with. Apple says excluding an app stops Siri from using what that app donates to personalize your experience. Whether that also keeps the app out of the graph, Apple doesn’t say.
• Look at your own graph. Pasquini’s open-source viewer iKnow runs locally and needs Full Disk Access for the Terminal you start it from. Take that permission away again when you’re done. Its optional AI-agent mode sends what it reads to wherever the model runs, so leave it off if you want the data to stay on the Mac.