Your Mac Is Quietly Building a Graph of Your Family: 51 Databases and XGBoost in macOS 26

You have probably noticed that Photos automatically titles albums "Mom" or suggests a "Family" memories set. It looks like magic, but there is very concrete infrastructure behind it. Researchers at 8rank have published an analysis of the undocumented knowledge graph that macOS builds on your computer, and it is one of the most detailed looks at how Apple does personalization without the cloud.

What they found inside

Starting with macOS 13, two daemons run on every Mac: knowledgeconstructiond and intelligenceplatformd. They store their results in the ~/Library/IntelligencePlatform directory. It contains 51 SQLite databases. The main one, graph.db, stores "subject — predicate — object" triplets drawn from a vocabulary of 181 classes and 278 predicates. It is a classic knowledge graph, only a personal one: the owner, their contacts, places, events, photos.

Of particular interest is how the system determines closeness between people. A pair of XGBoost classifiers look at 144 features. These are mostly behavioral signals: whom you message and how often, whom you appear with in photos, and on which day of the week and at what hour contacts happen. The first model answers the question "family or not," while the second separates "family or friends" from everyone else. In other words, there is no deep semantics: just ordinary tabular features and gradient boosting, which makes sense for a device, since the model is cheap and works offline.

What remains unclear

Another fifteen columns, from "mother" to "colleague," are filled in by some mechanism the authors could not trace. It may be a separate model, it may be rules, or data from other subsystems. In fairness, the researchers do not have the full picture.

The second important detail concerns deletion. When a relationship is revoked, the row does not disappear but moves to a separate table, retaining the model's confidence and the sources from which the conclusion was drawn. For the user, this means that a "forgotten" contact may leave a trace in local storage.

What to do about it

For macOS developers the takeaway is simple: the device holds a rich layer of user context, and Apple's system features rely on it. If you are building a mail, photos, or contacts app, it is worth understanding what expectations this graph creates for users and what data the system has already inferred on its own.

For product managers and founders, this is a lesson in privacy design. Local processing is not the same as transparency: the inference of who is whose mother is made without explicit consent and without an interface for review. If you build similar personalization, give users a way to see and correct the inferences, and when deleting, erase derived data too, including confidence scores.

For security specialists and corporate admins, this is a new surface to audit. The databases sit in the user library, which means they end up in backups and forensic extractions. Sensitive profiles, such as those of journalists or lawyers, should be evaluated from this angle.

Finally, you can view your own graph: 8rank released a local dashboard, iKnow, that reads the databases on your Mac without sending data anywhere. Run it on your own machine and check what the system thinks about you. It is useful to compare the result with reality: classifier errors show how fragile such inferences are if you build anything important on them.

The research applies to macOS 26, so behavior in other versions may differ. But the very fact that such a graph has been built for years without explicit settings deserves the attention of everyone who works with user data.