ada

Ada is a local-first, privacy-first personal AI assistant project named after Ada Lovelace. It aims to feel like a capable personal companion with natural voice interaction, persistent user-controlled memory, contextual awareness, and useful computer control, while keeping privacy, transparency, and human control as first-class requirements. The architecture is designed so that the AI model is not the final authority: permissions are deterministic, with AdaGuard using the Cedar policy engine to decide whether proposed actions are allowed. Learning does not silently create new rights, and private data is not automatically shareable. Memory is user-controlled, file-native, and Markdown-first, with current human-editable files as authoritative. The project follows principles such as local first, cloud optional and explicit, privacy by architecture, least privilege, and evidence before decisions. It reuses existing maintained systems where possible, including PydanticAI for agent runtime, Cedar for permissions, DBOS for durable external actions, and Ollama for local model serving. The current status is an initial implementation foundation, with some protections implemented and others being implemented incrementally.