Apple has a long-standing reputation for doing AI differently — prioritising on-device processing over cloud convenience. The new memory layer in Apple Intelligence represents the maturation of that philosophy into something genuinely useful.
How It Works
Rather than shipping your conversations to a server, Apple's on-device model builds a compressed semantic graph of your preferences, frequently visited places, communication style, and calendar patterns. This graph is encrypted with your device key and stored in the Secure Enclave — Apple's servers never see it.
When you ask Siri to "book somewhere nice for my anniversary dinner," the model already knows your preferred cuisine from past messages, the neighbourhood you frequent, and your typical spending range — without you having to spell it out.
The Differential Privacy Angle
For aggregate improvements (training the model to be better at certain tasks), Apple applies differential privacy — a technique that adds mathematical noise to any data that leaves the device, making it impossible to reverse-engineer individual behaviour even from population-level statistics.
"We're not building a profile on you. We're building a model that only you can use." — Apple VP of AI, WWDC 2026.
What Developers Can Access
The memory API will be available to third-party developers under strict App Store guidelines: apps can read from the memory graph only with explicit user permission per-query, and they cannot write to it directly. Apple controls what gets remembered and when.
