The privacy layer for the AI era

Use the world's best AI.Even for confidential work.

Keep using ChatGPT, Claude, Gemini, or any LLM. Amnesia AI® changes what they see, not how you work.

See how it works

The same chat you already know. The only new habit is opening Amnesia AI®.

How it works

Same AI. The difference is what it sees.

Amnesia AI® masks sensitive details on your device before anything reaches ChatGPT, Claude, or Gemini. Real answers come back, without compromising confidentiality.

01
You type normally

Real names, real questions, real documents, like any chat.

…conditions for Ridgeline?
02
Anonymized before the LLM

On your device, identifiers become coded stand-ins the AI can see.

RidgelineCompany_2
03
Originals restored

The reply uses stand-ins. Amnesia AI® restores the real names, locally.

Company_2Ridgeline
Zero-trust by design

Privacy has to be built in, not promised.

Masking and restoring run on your device, so not even we can see your data.

"We could be required to produce that, and I think that's very screwed up."
- Sam Altman, CEO of OpenAI
NYT v. OpenAI (2025) ordered every ChatGPT log preserved, even deleted chats. US v. Heppner (2026): terms of service alone erased any expectation of privacy.
Runs locallyReview-first modeHash-chained receiptsFail-closedPatent pendingRead the full security architecture →
Use cases

Built for work where a leak is a crime

Tax
IRC §7216

One unauthorized disclosure is a federal crime with personal liability.

Legal
Attorney-client privilege

Draft, summarize and research on client matters without handing them to a third party.

Health
HIPAA

Patient identifiers never reach the model. Built for data-minimization workflows.

Finance
GLBA

Client accounts and holdings stay on your machine. The AI reasons about stand-ins.

Amnesia AI® does not create attorney-client privilege or work-product protection, and does not determine whether information is discoverable. Built for professionals bound by GDPR, HIPAA and IRC §7216, designed to support data-minimization workflows.

More than confidentiality

Privacy was the requirement. This is the upside.

Auto-routing
Never overpay for a simple question.

Every message is scored for difficulty in under a millisecond, then sent to the cheapest model that can actually handle it: simple asks skip the flagship-model price tag, hard ones still get the frontier model.

Client memory
It remembers your clients. No one else finds out who they are.

Amnesia AI® keeps an encrypted, per-client memory on your device, so the assistant recalls context across sessions and years like a long-time colleague would. The provider only ever sees the masked version.

Protect your first confidential prompt.

No manual redaction. No weaker models. No separate workflow.

Available now for macOS & Windows

Five questions everyone asks

Does masking make the AI dumber?

No. Money, dates and context are preserved, and only identifiers are swapped for consistent stand-ins. Same models, full quality.

What if detection misses something?

Review-first mode shows you the exact payload before send; you can mask anything manually; and when uncertain, Amnesia AI® blocks the send entirely (fail-closed).

What do Amnesia AI®'s servers see?

Masked text and token counts. Content logging is architecturally impossible, not just prohibited by policy.

Do I need a powerful computer to run Amnesia AI®?

No. Everything is heavily optimized to run on a normal laptop. That is the whole point of Amnesia AI®: you get the privacy of a local model with the full power of frontier models.

Can I prove what was sent if I am audited?

Every send produces a tamper-evident, hash-chained receipt. One click verifies the whole chain, so you can show exactly what left your device.

All questions →