> ## Documentation Index
> Fetch the complete documentation index at: https://docs.asklethe.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Privacy Model & TEE

> The long-term privacy vision, and what's actually live today.

<Warning>
  This page describes a planned direction, not shipped functionality. See [How LETHE Works](/how-it-works) for the current architecture, and [Why LETHE Exists](/why-lethe) for the full case behind this design.
</Warning>

## The vision

Inference you don't have to trust an operator with. Your prompt is encrypted client-side before it ever leaves your device, then processed inside a Trusted Execution Environment (TEE): an isolated enclave where the hardware itself prevents anyone, including LETHE's own infrastructure and staff, from reading what's inside while it runs. When the request finishes, a cryptographic attestation proves it ran in that enclave and was discarded afterward. Not a policy that says the data was deleted. Proof.

That's the difference between "we promise not to look" and "we structurally cannot look." Candidate TEE providers under consideration: AWS Nitro Enclaves, Google Confidential Space, Phala, and Marlin.

## What's actually live today

* API requests (`/v1/*`) are stateless and not stored server-side, but they are **not** processed inside a TEE, and there is no attestation of deletion yet.
* Signed-in chat and Studio history is stored as regular, unencrypted rows in Supabase Postgres. This is the known gap between the vision above and where the product is right now.
* Guest history stays in browser `localStorage` only, private by virtue of never leaving your device, not by any server-side encryption guarantee.

Guest mode is currently the mode that most closely matches the "nothing remembered" promise end to end, precisely because nothing is ever sent to a server-side store in the first place.

## Planned mitigation

Client-side encryption of signed-in history, or moving that storage behind a TEE-backed store, is the leading candidate fix for the plaintext-Postgres gap. Not yet built. This page will be updated the moment it ships, not before.