HeartbeatHocusFocusFilesMacHyperTile
Guide

knowledgeconstructiond, the daemon that turns your Mac into a knowledge graph

knowledgeconstructiond reads your messages, photos and contacts to build the graph behind Apple Intelligence. Measured: five runs in 29 days, charger only.

By Chethan Kumar S N · Updated

Plug in your MacBook, walk away, and come back to find the fans spinning over a process with an unpronounceable name. knowledgeconstructiond, Activity Monitor says, and the percentage is real. Unplug the charger and it vanishes, which is the most confusing part. It is not broken. It is waiting for exactly this moment.

The name decodes into three moves. Knowledge, as in a knowledge graph, a map of how your things connect. Construction, as in building that map. And the d, the family suffix, because this is a daemon, a background process launchd starts without asking you. What it builds is the private, on-device index that Apple Intelligence and Siri Suggestions stand on: who you talk to, what your photos contain, where you go, what your apps are for, woven together so the system can guess what you need next.

Apple’s own manual page, knowledgeconstructiond(8), is three sentences and dated June 6 2022: a daemon that analyzes content on the device to build a general purpose knowledge graph, working together with intelligenceplatformd, performing some of the heavy processing tasks. The date matters. This stack is older than Apple Intelligence, which arrived two years later. The graph grew out of the Siri Suggestions era, and the older names are still in its wiring, as you will see below.

Here is what it looks like on a real Mac, macOS 27.0.1 on a MacBook Pro, 29 days of uptime, the same machine this site’s other daemon pages are measured on.

The receipts

WhatMeasured
Path/System/Library/PrivateFrameworks/IntelligencePlatformCore.framework/Versions/A/knowledgeconstructiond
Size266,496 bytes, three architectures in one file: x86_64, arm64e and arm64e.x1, so it still ships for Intel Macs
SignatureApple platform binary, identifier com.apple.knowledgeconstructiond, on the sealed system volume
UserThe logged in user, not root
Parentlaunchd, process 1, always
Runs5 since boot, 29 days ago, minimum runtime 10 seconds per run
Last exitJETSAM_REASON_MEMORY_PERPROCESSLIMIT, the system’s per-process memory cap
Partnerintelligenceplatformd, alive at 13,280 KB of memory with 0.14 seconds of processor time in 47 minutes on the clock
Manual pageknowledgeconstructiond(8), dated June 6 2022, see also intelligenceplatformd(8)
Log volumeZero lines in a three hour window: between runs, it does not exist
Crash recordOne knowledgeconstructiond entry in CrashReporter, dated October 11

Two rows are the story. Five runs in 29 days is not a malfunction, it is the schedule: the main job only runs on external power. And the last exit reason says the system killed it for memory, which sounds alarming and is simply how macOS polices a job flagged ResourceIntensive: work hard, hit the cap, get recycled, resume next window. The binary itself is a thin front, 53 strings in the whole file, a launcher whose heavy code lives in the framework libraries around it, which is exactly what the manual page means by performing some of the heavy processing tasks.

The search picture matches a name people meet by accident. The bare seed returns four suggestions, all of them this daemon: knowledgeconstructiond, knowledgeconstructiond process, knowledgeconstructiond meaning, and com apple knowledgeconstructiond. The what is family tops out at what is knowledgeconstructiond on mac, the exact question this page answers. There is no high cpu row and no virus row, and the spelling people type breaks the name apart, knowledge construction d mac, because nobody knows where the word ends.

What the job board says

The launch configuration is the best description anyone will write of this daemon, because Apple named its three scheduled activities in plain English.

Activity, Apple’s own nameScheduleThe rules
IntelligencePlatformCore.FullPipelineEvery 86400 seconds, dailyRequiresExternalPower, PowerNap, Maintenance priority, ResourceIntensive
IntelligencePlatformCore.AutonamingMessagesPipelineDailyCPUIntensive, AllowBattery false, PowerNap
IntelligencePlatformCore.AutonamingMessagesAggregationAccuracyEvery 604800 seconds, weeklyCPUIntensive, AllowBattery false, an accuracy audit

Read the middle row again, because it is the rare case of a private daemon naming the exact feature it serves. Autonaming Messages is the Apple Intelligence feature that suggests names for group conversations, the reason a chat about ski tickets quietly became Ski Trip in your sidebar. For that suggestion to exist, something has to read the conversation, figure out who is in it and what it is about, and write the answer down where the Messages app can find it. That something is this daemon, daily, while the battery is safe at home.

There is a fourth entry, and it explains the weirdest timing detail. A one-shot task called Pipeline.FastPass is gated on a dependency with a completion notice from com.apple.cloudphotod.initialdownload.fastpass, the first full download of your photo library by the iCloud Photos daemon. In plain English: before the knowledge pipeline does its first pass, the Mac waits until your photos have fully arrived. A half-downloaded library would build a half-true graph, so the system orders the work: photos first, understanding second.

And that is the whole mystery of the vanishing daemon. The main pipeline carries RequiresExternalPower, the autonaming jobs carry AllowBattery false. On a MacBook that commutes on battery, this process may not run for days. Plug in overnight and it catches up, all at once, which is precisely when people notice it.

The model shelf

The framework the daemon lives in carries its own library of tiny on-device models, forty five items on disk, and their filenames read like a job list for understanding a life:

Model on diskWhat it is for
EntityTagging_FamilyAndFriends.mlmodelcTelling family from friends in your data, 36 KB
EntityRelevanceModel.mlmodelcRanking which people and topics matter to you
EntitySimilarityLinearModel.mlmodelcDeciding two mentions are the same person or thing
DiningOutModel.mlmodelcWorking out that you are at a restaurant
LOITypeToOneHotTransformer.mlmodelcClassifying locations, the routine places you go
Top10KPopularBundleIDs.trieA lookup trie of the ten thousand most popular apps, 241,668 bytes
WalletTransactionOrder.srccfgMaking sense of wallet transactions

DiningOutModel is the one to sit with. There is a model, on your disk, whose only job is to conclude that you are eating out. Multiply it by the rest of the shelf, the wallet configs, the ScreenTime config, the Relationship config, and you can watch the knowledge graph being assembled in advance: places, people, money, time, apps, each with its own small classifier, none of it a large language model, all of it running on the Neural Engine.

The anatomy of a reader

The sandbox profile, com.apple.knowledgeconstructiond.sb on the sealed system volume, is 183 lines long and stamped Copyright 2021, three years before Apple Intelligence was announced. Deny by default, then a careful list of what a knowledge graph builder needs. Read the list and you have read the job description.

It reads the photo knowledge graph directly, photosgraph.kgdb inside your Photos library, the database the photo analysis daemon fills with what your pictures contain. Read only, no writing back. It reads the people database at /private/var/db/CoreDuet/People/interactionC.db, the system’s record of who you actually interact with, and may write the bookkeeping files beside it. It is a full contacts client. It reads preference domains for Maps, Calendar and Photos, and a folder called /Library/Trial, the system’s feature-flag stash, to know which experiments it is allowed to run.

Then the phone book of helpers it may call: the calendar data daemon, the reminders daemon, the Calendar agent, both faces of the suggestions daemon, the routines daemon that knows your regular places, the family sharing agent, the home kit daemon, the location daemon, the media analysis service, the accounts manager, the permissions broker, Spotlight’s indexing agents and the model manager that serves on-device models. Two entitlements close the picture: HealthKit and HomeKit read access, and a direct client for the Apple Neural Engine.

Now the detail that answers the question everyone actually has. In all 183 lines there is no network permission. Not one outbound rule. The daemon that reads your messages, photos, contacts, health and home data is structurally incapable of sending any of it anywhere. The graph it builds lands in ~/Library/IntelligencePlatform, and when an administrator’s terminal tries to look inside, macOS answers Operation not permitted, the same privacy wall that guards the account and calendar stores in this series. On the measured Mac even the directory listing bounces.

The honest cost

Measured plainly: five runs in 29 days, zero log lines in the window I checked, and a partner process that has barely registered on the processor at all. This is not a daemon in your CPU ledger, because it is not in your CPU at all unless the charger is attached. What people experience as knowledgeconstructiond being greedy is one sitting of catch-up work, memory-capped and recycled by design, with the crash report to prove the system was watching. If it never ran, the cost would be invisible, and so would the feature: group chats that name themselves, suggestions that know who you mean. The work is the point; the schedule is just polite about it.

Questions

What is knowledgeconstructiond on my Mac?

It is the process Apple's own manual page calls a daemon that analyzes content on the device to build a general purpose knowledge graph. It lives inside the IntelligencePlatformCore framework, the plumbing under Apple Intelligence, and it works together with a partner process called intelligenceplatformd, doing the heavy processing while the partner coordinates. Every so often, when the Mac is plugged in and quiet, it wakes up and reads what the system already knows, your messages, photos, contacts, calendar entries and locations, and weaves it into a private map of your data that features like suggested names for group chats are built on. It is launched by launchd, runs as your own user, is signed by Apple on the sealed system volume, and has a manual page, knowledgeconstructiond(8), dated June 2022.

Why is knowledgeconstructiond using so much memory?

Because it is doing model work in a hurry and the system caps it on purpose. The daemon's last exit on the measured Mac was JETSAM_REASON_MEMORY_PERPROCESSLIMIT, which means macOS killed it for crossing its per-process memory limit, and there is a crash report record for it on disk from October. This is the system's pressure valve, not a fault: the job is flagged ResourceIntensive, it is allowed to eat memory while it runs, and when it crosses the line the system recycles it and the work resumes next window. A memory spike while the charger is connected is the daemon doing its day's work in one sitting, not a leak.

Is knowledgeconstructiond reading my messages?

Yes, in the specific way its configuration says out loud. One of its scheduled jobs is named AutonamingMessagesPipeline, the machinery that suggests names for group conversations, and it cannot do that without reading the conversations. It also holds read access to the photo knowledge graph, the people interaction database, your contacts, and preference data from Maps, Calendar and Photos. Two things keep that honest. The graph it builds sits in a sealed store that even an administrator's terminal gets Operation not permitted on. And the sandbox profile grants it no network access at all, deny by default with no outbound rule, so the analysis itself has no way to send anything anywhere. The reading is real, and it stays on the device.

What is the difference between knowledgeconstructiond and intelligenceplatformd?

They are a pair, and the manual page introduces them together: intelligenceplatformd coordinates, knowledgeconstructiond does the heavy lifting. In live measurements the coordinator sits in memory all day at roughly thirteen megabytes and has barely any processor time to its name, while the construction daemon is absent entirely until its window opens, then works hard and disappears. There are older relatives too: siriknowledged, the Siri-era knowledge daemon, and spotlightknowledged, which feeds search. If you see knowledgeconstructiond chewing while intelligenceplatformd idles beside it, that is the pair division of labor working as designed.

Why do I see knowledgeconstructiond more than once?

Check whether Xcode simulators are running. Every iOS simulator runtime ships its own copy of the IntelligencePlatformCore framework inside its cryptex mount, and a booted simulator starts the platform the way a real iPhone would. On the measured Mac two extra copies were alive at once, from the iOS 27.0 and 27.1 runtimes, running from paths under /private/var/run/com.apple.security.cryptexd. The same simulator-twin pattern shows up across this series, on the calendar sync daemon and the Family Sharing daemon among others. Quit the simulator and the extras go with it.

Can I quit or disable knowledgeconstructiond?

There is nothing to configure and nothing to gain by quitting it. The manual page documents no options, launchd starts it on schedule rather than keeping it alive, and on the measured Mac it had run only five times in 29 days, because its main job waits for external power. Force quitting buys you nothing but a half-built graph, since the next charging session picks the work up again. If your goal is less of this machinery entirely, the lever is upstream: it only works with what the Apple Intelligence platform has already collected, so the features and data sources you turn off in Settings are the real controls. The process itself is just the workman.

Ready? Buy Ribbon for $15, or read all the guides.