Your Mac keeps a family file on you, and Apple won't explain it

What this is about

Your Mac has been running something for years that doesn't show up anywhere in System Settings: a personal knowledge graph. This is a database that doesn't store individual files but relationships between things, such as “Person A is Person B's sister” or “Event X takes place at Location Y.” It's built by two background services called knowledgeconstructiond and intelligenceplatformd. Apple describes them only in a terse man page from 2022, the short technical help page you can pull up in Terminal. It says the services analyze content on the device to build a knowledge graph that “represents the device and the user.” Other parts of the operating system can query it to provide “richer experiences.” That's all Apple says.

Security researchers Dario Pasquini and Michał Bazyli, publishing as 8rank, have now taken this store apart. Their analysis, “Every Mac comes with an (undocumented) family member classifier,” is currently making the rounds on Hacker News. At the same time, Apple is expanding how heavily it uses personal context with Siri AI, Live Rewind and Siri Recap. This article digs into three questions: What's in the graph? What's observed and what's merely inferred? And who gets to see it?

The databases at a glance

According to 8rank, ~/Library/IntelligencePlatform holds 51 SQLite databases, stored unencrypted. The only thing protecting them is the macOS Full Disk Access permission, which you grant to individual apps in System Settings. On top of that come older activity stores that forensic investigators have known about for years.

The graph itself is built on an ontology that Apple ships with the system, a fixed catalog of concepts with 181 classes and 278 predicates. These include 57 activity classes (sleeping, commuting, exercising, eating, making phone calls) as well as categories for orders and payments. Back in 2023, a Hacker News user found predicates for religious affiliation and political party membership in ontology.db. But a class existing doesn't mean it actually gets filled, as the 8rank authors also stress.

As sources, 8rank lists Contacts, Calendar (titles and attendees), Photos (faces, and who appears in a picture with whom) and the metadata of Messages and Mail: who writes to whom and when, but not the content. Add to that Family Sharing including members' ages, the emergency contact from Medical ID, Screen Time, Siri Suggestions, installed apps and location history.

The databases are real, as independent forensic findings show. Alexis Brignoni's forensic tool iLEAPP has been parsing graph.db on iPhones since a merge on September 23, 2026. According to that work, the file exists from iOS 16 onward and is populated from iOS 17. The Mac counterpart DLEAPP followed on October 30, 2026, tested against a macOS 15 image. The iLEAPP developers add an important caveat: an entry doesn't prove that the user created it, confirmed it or actually experienced the event.

The family classifier and what “confidence” really means

The most striking find: the system contains two CoreML models, Apple's format for machine learning models that run locally. They're called EntityTagging_Family and EntityTagging_FamilyAndFriends. Both are XGBoost classifiers, a classic technique built from many small decision trees. They work with 144 features: around 20 on identity (same last name, same address, Family Sharing, age gap), roughly 40 on communication (when and from where you message someone) and roughly 84 on photos (who shows up in pictures with you, when and where). According to 8rank, the Photos app also runs twelve analyses of its own that estimate partners, parents, children, colleagues and friends. If Apple Intelligence is turned on, a local language model can additionally match names and relationships to faces, using the text of your messages to do so.

This is where you have to separate carefully what's proven from what isn't:

• Proven, according to 8rank, is that the models, the feature list and the 17-column table exist.
• Unclear is how the other 15 columns are calculated. Apple ships a model only for “family” and “familyAndFriends.” It's also unclear how the probabilities get turned into confidence values in the graph.
• Coarse describes the values themselves: in the real table, the authors found only a handful of levels, namely 0.125, 0.25, 0.5 and 0.96. Their nicely graded sample charts are based on made-up test data.
• Misleading is the term “confidence.” According to the iKnow manual, the value in the graph usually indicates which source wrote the entry, not how certain the statement is: Contacts yields 1.0, Photos 0.9, most other sources 0.7 to 0.8. A high family score on its own doesn't prove a family relationship.

There's more: the authors explicitly label their attribution of mechanisms as inference. Many entries carry the note “could not find mechanism,” meaning they couldn't work out how the value comes about. By our reading, the “where you'll be next” view in the iKnow tool is also essentially based on calendar data (time, room, attendees, dial-in details). It's not a dedicated AI predicting your next location. Sleep times are derived too: in the example, message timestamps produced a quiet window from 1 to 7 a.m. Nobody stored that as “sleep”; iKnow calculates it.

iKnow: check for yourself, with caution

The iKnow tool is a local web interface written in Python. It opens the databases read-only, runs exclusively on 127.0.0.1 and needs Full Disk Access for Terminal. It's tested on macOS 26.6.2 and 26.7.1, and only partially on 27.0.1. The README itself admits the tool is “fully LLM-generated and may not be 100% accurate.” The repository currently has two commits. We weren't able to run iKnow on a Mac ourselves. Our information is based on the README, the agent manual and the 8rank analysis.

The manual for AI agents is worth noting. Before an agent reads anything, it has to ask, word for word, whether the user trusts their LLM provider. The reason: a coding agent with shell access in a terminal that has Full Disk Access could read the graph and send it to its cloud provider. Our take: if you run agents like Claude Code or Codex in a terminal with full access, keep this in mind.

What Siri AI, Private Cloud Compute and Google get to see

In the man page, Apple writes that other parts of the system query the graph via intelligenceplatformd. It doesn't say which ones. 8rank found copies for individual consumers, including a Siri folder, and cautiously writes that the store is one from which language models “reportedly” retrieve information. The authors offer no evidence for that.

Officially, Apple describes Siri AI's “Personal Context” features differently: via a semantic index built on Spotlight, a meaning-based search across Mail, Messages, Photos, Notes and Calendar, and via App Intents, the interface apps use to provide data. Apple says this search runs on the device. According to Apple Support, Siri stores relationships like “Ashley is my mother” encrypted in Contacts, which syncs via iCloud. Contact relationships, in turn, are one of the graph's input sources. Whether Siri AI accesses graph.db or the family scorecard directly isn't publicly documented.

Anything not handled on the device goes to Private Cloud Compute (PCC), Apple's hardened servers. Apple promises stateless processing there: data is used only for the request and is neither stored nor accessible to Apple. Since WWDC in June 2026, however, PCC has been running partly on Google Cloud hardware with Nvidia Confidential Computing, Intel TDX and Google's Titan chip. According to Apple, this applies to “agentic tool use and complex reasoning.” MacRumors and Help Net Security quote Apple admitting that PCC on Google Cloud was “not fully implemented” during the beta. Gemini serves as the foundation for Apple's Foundation Models. Apple stresses that user data isn't processed directly on Google's servers. What it apparently means is Apple's software on Google's hardware. Apple doesn't disclose in detail which parts of your personal context get sent there with each request.

Siri Recap on the Apple Watch, which we covered recently, is a concrete example. Alongside the condensed conversation text, the feature sends calendar events and rough location labels like “Home” or “Work” to PCC. That shows derived context does leave the device in small portions, albeit for a specific purpose.

Important for European readers: Siri AI is blocked in the EU on iPhone and iPad for now because of the Digital Markets Act. On macOS 27, though, it's available. Apple justifies the block by saying that, as the EU reads the rules, it would have to give “any virtual assistant direct access to users' private data.” That's awkward, because a good deal of that data sits in exactly these kinds of stores.

Can you turn it off?

Short answer: not in any documented way.

• The graph predates Apple Intelligence (2022 man page, macOS 13). According to Apple Support, turning off Apple Intelligence removes the local models. But CoreGuard points out that this doesn't guarantee all the knowledge services go away. And an HN user who's staying “on Sequoia for now” isn't off the hook either: macOS 15 is affected.
• According to Engadget, “Learn from this App” only controls app suggestions. But according to 8rank, Siri Suggestions feed into the graph. Nobody has shown whether turning off suggestions removes data from the graph. For Siri AI, Engadget recommends disabling “Show Content in Search” on a per-app basis.
• If you delete graph.db, it gets rebuilt. That's according to an Apple Community thread in which the file had grown to over 4 GB. On another Mac it was 3.5 MB. CoreGuard considers scripts that disable System Integrity Protection (SIP) for this purpose both risky and ineffective.
• Deleted contacts disappear from the active graph, but the entries move into expired_stable_graph along with their confidence value and source. According to 8rank, they're supposed to stay there for about 48 hours, but the authors found very old leftovers. They can only be removed by hand.

We found no statement from privacy groups like noyb or European consumer organizations specifically addressing this on-device profiling.

Our take

Apple personalizing your experience isn't new, and on-device processing is a real advantage over the cloud profiles kept by Google or Meta. What's new is how structured the result is. Raw data like photos, chats and calendars has always lived on your Mac. The graph turns it into ready-made answers: who's family, who's close, when someone sleeps. In their second piece, “Weaponizing Personalization,” Pasquini and Bazyli show that a language model could derive an extortion plan from the graph alone: targets from the relationship scores, timing from the calendar. Their way of putting it: the knowledge graph “does the work in advance.” That does require code execution with Full Disk Access on the target Mac, though. So this isn't a remote attack but a post-compromise scenario.

Our assessment: the critics' strongest argument isn't the classifier itself but the complete lack of transparency. There's no view in System Settings, no way to correct anything and no documentation of which services read the data. It also affects third parties who were never asked: parents, friends, colleagues. At the same time, many of the detailed claims are shaky. The mechanisms are inferred, the tool was written by an LLM, and the confidence values mean less than their name suggests. Above all, the most important question remains open: whether and how Siri AI uses this graph, and how much of it ends up in Private Cloud Compute, which now runs partly on Google hardware. Only Apple can answer that.