Savant / Privacy & control
YOUR CONTEXT. YOUR AUTHORITY.
Remembering is a choice.
So is forgetting.
Savant’s ownership goals include memory you can inspect, correct and remove, and clear authority over accounts, tools and cloud use. Product requirements and current website behavior are separated below.
Local inference.
Clear boundaries.
Running a model locally does not make every connected service work offline. Model downloads, email synchronization, weather and some device services may still need an internet connection.
Private context should not silently become a cloud request. The product design requires explicit cloud choice and relevant permissions. This website does not establish that a complete runtime permission gate has shipped.
The walkthrough, hardware planner and guide run in your browser. The planner uses only what you type; it does not inspect your hardware. These tools do not send their inputs to a model, account or device, and do not save them in browser storage. Ordinary page and asset requests still reach the website host.
PLANNED MEMORY CONTROLS
You should be able to understand
what the system remembers.
Inspect and correct.
The intended experience should let you see a stored preference or record, understand why it is relevant, and correct it when circumstances change.
A remembered statement should not become verified truth simply because it was saved.
Remove and limit.
Keeping context should remain an owner choice. Removal and retention controls need clear behavior, including how a removal affects later retrieval.
These are design requirements awaiting their own product evidence, not controls that this website exercises over personal memory.
ACCESS IS NOT IMPLIED
Accounts and actions
need their own authority.
An email connection requires access to the intended account. Reading a message, drafting a response and sending it are distinct actions. Likewise, describing a home routine does not establish device support or approval to execute it.
The eventual integration should make the requested action, permission and observed result clear. A model’s confident answer is not proof that an action happened.
KEEP THE LEARNING HONEST
Context is not model weights.
Guided memory setup is intended to collect useful preferences and context. It does not mean every conversation trains or fine-tunes the foundation model.
Corrections and verified outcomes should be retrieved when relevant and checked for usefulness. Learning from experience still needs evidence that the next result improved.
The guide is information, not a connected agent.
The persistent Site guide searches authored website guidance. It is not connected to an AI model, mailbox, payment account or device. For release status, use the Development & results page.