How Siri Uses Message History in iOS 18

In iOS 18, Apple Intelligence equips Siri with personal context awareness, allowing the assistant to retrieve specific details hidden within your historical text messages. By shifting from simple keyword queries to an advanced, on-device semantic retrieval system, Siri can understand conversational requests, locate relevant message threads, extract precise information, and present a direct answer while preserving user privacy.

On-Device Semantic Indexing

At the foundation of this capability is an on-device semantic index. Instead of merely matching exact words like older versions of Spotlight search, iOS 18 utilizes small, localized machine learning models to map the meaning of your conversations. As you send and receive texts, the system processes names, dates, locations, links, and shared media into mathematical representations called embeddings. This allows Siri to understand context—such as recognizing that "the Italian place" in a text refers to a restaurant recommendation.

The Personal Context Engine

Apple Intelligence introduces a dedicated Personal Context engine designed to bridge data silos across native iOS applications. When a user asks a question—such as "What time does my brother's flight land?"—the engine recognizes that the answer likely resides in recent communications. It queries the local message index to trace messages from contacts flagged as family, analyzes timestamps, and isolates flight numbers or arrival estimates without requiring the user to specify which app holds the information.

Local Retrieval-Augmented Generation (RAG)

Once Siri identifies potential message threads that answer the query, it utilizes a retrieval-augmented generation (RAG) pipeline:

  1. Query Parsing: The on-device Large Language Model (LLM) deconstructs the user's spoken or typed prompt to determine intent and required parameters.
  2. Context Retrieval: The semantic engine retrieves the most relevant snippets of past message exchanges related to those parameters.
  3. Synthesis: The local model synthesizes the retrieved text snippets and generates a concise, natural response directly addressing the prompt.

Privacy and Data Security

All indexing and basic retrieval processes for historical messages occur entirely on the user's device via the Apple Neural Engine. Text history is not uploaded to external servers to build advertising profiles or train global models. In scenarios where a request exceeds the computational limits of the on-device model, Siri utilizes Private Cloud Compute. In these rare instances, only the exact data required to fulfill the specific prompt is sent to secure Apple silicon servers, processed without permanent storage, and returned immediately to the device.