Apple Siri memory system in iOS 27.3 code adds 60 questions

Early code findings indicate an Apple Siri memory upgrade is underway, centered around a newly uncovered personal memory system spotted in an iOS 27.3 test build. Uncovered by code researcher pdfu, the architecture aims to map daily routines, relationships, and personal goals, giving Apple's assistant long-term contextual awareness rather than just answering one-off requests.

Targeted at users who want Siri to act like a proactive personal assistant, the system uses on-device models and Private Cloud Compute to organize facts and health patterns. However, the framework remains incomplete, with no official release date and no visible settings yet for managing user data.

Inside the Memory Graph and Memory Atoms

The leaked code reveals two structural foundations designed to store personal details without losing context. At the center is the Memory Graph, an organizational layer that tracks relationships, frequent locations, daily routines, fitness habits, and the factors that either motivate or hinder a user from achieving their goals.

Populating this graph are smaller data units labeled Memory Atoms. These atomic entries record specific facts, opinions, intentions, and events, while tethering each detail to corresponding people, places, timestamps, and topics so Siri can connect disparate pieces of information over time.

Dual Processing Engines and 60 Diagnostic Prompts

According to the code analysis, the architecture relies on two separate ingest pipelines to build this profile. The first mechanism translates structured system data into actionable insights, such as parsing name fields directly from a user's contact card. The second pipeline utilizes on-device machine learning models alongside Apple's Private Cloud Compute infrastructure to distill context from Siri conversations, Messages, Photos, and Health records.

Currently, only one AI-driven evaluation mechanism is visible in the code. This component appears designed to uncover user motivations, obstacles, and available resources using a suite of 60 built-in diagnostic questions, such as identifying specific daily routines that might interfere with getting adequate sleep.

Kùzu Database Integration and Missing Controls

Under the hood, the infrastructure relies on an Apple-adapted version of Kùzu, an embeddable graph database technology Apple acquired in 2025. The code also points to a distinct, explicit memory storage partition set aside for search queries, though no active processes are currently writing data to that storage area.

Despite the functional architecture shown in the files, the memory engine is still in an early development phase. Key operational pipelines, including the automated generation of Memory Atoms, are not yet functional in the test build. While existing versions of Siri already scan emails and photos for calendar context, this deeper memory system currently offers no timeline for deployment and lacks any exposed user-facing privacy controls.

Would you prefer Siri to actively log your habits and sleep roadblocks to provide proactive coaching, or should digital assistants stick strictly to handling requests on demand?