Siri Memory Appears to Use a Hybrid Index and Graph

What was found in Siri’s memory system

I would not call this a finished public memory system, but the direction is fairly clear: Apple appears to link personal context with indexes and structured entities. At the time of this analysis, the company had not published a complete technical specification for Siri’s long-term memory.

Apple’s Spotlight and App Intents documentation describes the official architectural layer. Apps can index entities, provide user actions, and connect data to screens so Siri and Apple Intelligence can locate content through semantic search. Foundation Models supports multi-step sessions, but its context is not presented as persistent cross-app memory.

The most interesting details came not from Apple documentation but from a Reddit post describing reverse engineering of macOS beta software. It outlines a hybrid index: BM25 full-text retrieval works alongside vector search, with results merged through reciprocal rank fusion, or RRF.

A memory graph centered on the user is also mentioned. Its nodes reportedly cover contacts, identity, relationships, places, habits, events, health, and priorities. So-called memory atoms represent facts, opinions, intentions, and events connected to time, location, entities, and topics.

For now, this is not a platform contract but evidence of an internal implementation. Component names, schemas, and storage may change, and there is no public data on retrieval latency, accuracy, index size, battery use, or the boundary between the device and Private Cloud Compute.

Why a hybrid index changes assistants

If the finding is correct, Siri gains more than a conversation history: it gets a mechanism for selectively restoring personal context. “Selectively” is the key word here. BM25 helps with exact matches, vectors capture meaning, and the graph links retrieved information to a specific person, place, or event.

For developers, this could mean richer data discovery without public access to Siri’s internal memory. An app describes its entities and actions, while the system layer decides what to retrieve for a request. That boundary reduces direct access to other data, but it does not eliminate errors in the assistant’s own inferences.

The first things I would test are not polished answers but the correction of false facts, removal of outdated relationships, and protection against exposing context in the wrong situation. Personal memory remains useful only until it confidently remembers something that never happened.

We previously examined Rust LocalGPT, a local assistant with persistent memory and an API. Its approach helps explain how personal data and context may be retained in systems such as Apple’s developing platform.