Your Mac has been building a graph of your family, and Siri's agents will read it
The memory layer for Apple's agents already exists, and it's on your disk. Security researchers at 8rank published a teardown of what macOS builds about you, and it was the biggest tech thread on Hacker News Saturday with 526 points and over 300 comments.
Two daemons, knowledgeconstructiond and intelligenceplatformd, have shipped on Macs since 2022. They keep 51 SQLite databases in ~/Library/IntelligencePlatform, in plaintext, behind Full Disk Access protection. The core one, graph.db, is a triple store with 181 entity classes and 278 predicates: subject, relationship, object. It's fed by Contacts, faces in Photos, Calendar titles and attendees, Family Sharing including every member's age, the Medical ID emergency contact, people and events Siri Suggestions finds in Mail and Messages, and data apps donate through App Intents. An XGBoost model plus rules infers how people are related to you. Fifty-seven of the 181 classes describe activities like sleeping, commuting and dining. All of this is the researchers' reverse-engineering on macOS 26.6.2, and they flag many classes as declared but possibly unused.
Apple did publish a 2023 paper on the techniques. What's missing is documentation of what's collected on a given Mac, and anything about the kinship inference. The man page is close to empty. 8rank open-sourced a local dashboard, iKnow, so you can look at your own graph.
Why this matters for agents: when Siri and App Intents agents answer "text my sister" or "when's my kid's recital", this is what they draw on. That's the good version of personal memory, on device, not in a cloud. The weak spot is the other side of the same coin. One plaintext store behind one permission gate holds a map of your family, health contact and routines. Any agent or app you grant Full Disk Access, which many coding agents ask for, can read the lot. 8rank calls it a privacy single point of failure, and that's fair.
Practical take: think twice before granting Full Disk Access to agent tools, and look at what's in that folder before you do.
Write-up: https://8rank.ai/game/apple-intelligence-personal-knowledge-graph
Dashboard: https://github.com/8rank/iKnow
← Back to all articles
Two daemons, knowledgeconstructiond and intelligenceplatformd, have shipped on Macs since 2022. They keep 51 SQLite databases in ~/Library/IntelligencePlatform, in plaintext, behind Full Disk Access protection. The core one, graph.db, is a triple store with 181 entity classes and 278 predicates: subject, relationship, object. It's fed by Contacts, faces in Photos, Calendar titles and attendees, Family Sharing including every member's age, the Medical ID emergency contact, people and events Siri Suggestions finds in Mail and Messages, and data apps donate through App Intents. An XGBoost model plus rules infers how people are related to you. Fifty-seven of the 181 classes describe activities like sleeping, commuting and dining. All of this is the researchers' reverse-engineering on macOS 26.6.2, and they flag many classes as declared but possibly unused.
Apple did publish a 2023 paper on the techniques. What's missing is documentation of what's collected on a given Mac, and anything about the kinship inference. The man page is close to empty. 8rank open-sourced a local dashboard, iKnow, so you can look at your own graph.
Why this matters for agents: when Siri and App Intents agents answer "text my sister" or "when's my kid's recital", this is what they draw on. That's the good version of personal memory, on device, not in a cloud. The weak spot is the other side of the same coin. One plaintext store behind one permission gate holds a map of your family, health contact and routines. Any agent or app you grant Full Disk Access, which many coding agents ask for, can read the lot. 8rank calls it a privacy single point of failure, and that's fair.
Practical take: think twice before granting Full Disk Access to agent tools, and look at what's in that folder before you do.
Write-up: https://8rank.ai/game/apple-intelligence-personal-knowledge-graph
Dashboard: https://github.com/8rank/iKnow
Comments