macOS Builds Undocumented Personal Knowledge Graph Since Version 13, Inferring Family Ties via On-Device ML

Since macOS 13 (2022), every Mac runs a pair of daemons whose sole purpose is to build a detailed graph about the user and people around them. The daemons knowledgeconstructiond and intelligenceplatformd analyze Contacts, Calendar, Photos, Messages, Mail, Family Sharing, Health emergency contacts, Screen Time, and a long list of behavioral signals to determine identities, relationships, locations, and activities.

All data is stored locally in ~/Library/IntelligencePlatform as 51 SQLite databases protected by the Full Disk Access TCC permission. The most significant database, graph.db, holds subject-predicate-object triples using an ontology of 181 entity classes and 278 predicates. Each triple records either a direct fact or a relationship edge, complete with confidence scores, source references, and timestamps.

Relationship edges are reified through 46 dedicated relationship-type classes so that each edge can carry its own confidence value and be revoked later. The system copies raw facts from system apps, applies rule-based joins, and runs multiple machine-learning models. Two CoreML XGBoost classifiers (EntityTagging_Family.mlmodelc and EntityTagging_FamilyAndFriends.mlmodelc) consume 144 numeric features covering identity signals, communication patterns, and co-appearance statistics from photos to output family and friend probabilities each day.

Additional models include embedding networks for entity resolution, face-clustering logic inside the Photos app, and a local LLM (when Apple Intelligence is enabled) that reads message text to suggest names and relationship labels for unidentified faces. The resulting inferences are exposed through the undocumented iKnow panel, allowing users to inspect the stored graph directly on their device.

While the graph is rebuilt daily and forgets deleted contacts, the underlying models and stored triples remain opaque to users and lack public documentation beyond a brief man page. Researchers note that the same on-device store is intended to supply context to future Apple Intelligence language models.