Local storage architecture for offline assistants
Structuring encrypted calendar entries and reminder vectors in an isolated on-device database.
Elena Rostova
Systems Architecture & On-Device ML
Building a genuinely private voice assistant requires eliminating server dependencies for daily operational queries. If calendar items, reminders, and personal notes are stored in remote clouds, privacy promises remain vulnerable to network leaks and policy changes.
Dimlit stores all user entities in a local, encrypted sqlite database compiled directly into the application sandbox. Queries execute against on-device indices, ensuring immediate lookup times even in airplane mode or remote areas without cellular connectivity.
Isolation and encryption
Data security relies on native operating system keyrings rather than custom authentication schemes. When the device is locked, database encryption keys are purged from volatile memory until authenticated via biometric sensors.
Local vector embeddings allow semantic search across past notes without transmitting plain text to remote inference servers. The on-device engine indexes reminder context locally, keeping personal history strictly within device hardware boundaries.