AI That Burns After Reading

"Privacy-first AI" is having a moment. It's on every homepage, right next to the padlock stock photo, and it can mean anything from "we encrypt in transit" to "we pinky promise." The phrase has been polished into meaninglessness.

So we gave ours a rule you can picture instead: every AI glance burns after reading.

If you've used Hearth's Unsent Vault, you already know the gesture. You write the letter you'll never send, you seal it, and you burn it — gone, on purpose, forever. We built the AI on the same ritual. It reads your words for exactly as long as it takes to help you, and then that glance goes into the fire. Hearth's AI reads your journal — that's its job — and if an app is going to read your diary, you deserve to know exactly when, exactly how much, and exactly what happens to the words afterward.

What "burns after reading" means, concretely

When you ask Hearth's AI a question — through Inkling, or from ChatGPT or Claude via the connector — here's the entire life of your words on the server side:

The specific entries needed to answer are decrypted, in memory, in that moment. They're read once, an answer comes back, and then everything decrypted burns. Dropped. Not written to disk, not cached for next time, not tucked into a log file, not kept as a "conversation history" on our servers, not used to train anything — not our models, not anyone's.

The plaintext of your entry exists for a heartbeat, in RAM, because you asked a question. Then it stops existing. The technical term is zero retention. Not "we retain it responsibly." Zero. Burned.

The AI provider on the other end runs under zero-retention terms too: no storage, no training on your content, full stop. We wouldn't wire your diary to anything that keeps a copy.

The transient decrypt path

Slightly more mechanically, for the curious. Your entries live on our servers as ciphertext — locked boxes. In the default mode, Hearth holds one sealed spare key. (Why the spare exists at all, and how to delete it, gets its own post.)

When a question arrives, the spare unwraps in memory, opens only the entries relevant to your question, hands their text to the model under those zero-retention terms, returns the answer to you, and releases everything. The key goes back to being sealed. The plaintext burns. The boxes were never open longer than the question took.

No step in that path has a "save" in it. That's not an accident of the current version; it's the design. We built the fire; your words only pass through it.

"Enforced by tests" — what that means

Here's the part we're proudest of, and the part that separates this from a promise.

Our codebase has a verification suite whose whole job is to check the privacy invariants: that no plaintext gets persisted, that nothing readable lands in a log, that the decrypt path stays transient. If some future engineer — tired, well-meaning, debugging at 2am — writes code that stores what it shouldn't, the tests fail and the change doesn't ship.

"We promise not to keep your data" depends on everyone's character forever. "The build breaks if we keep your data" depends on nothing. We'd rather be policed by our own test suite than trusted on vibes.

The daily budget

One more guardrail, because "the AI can read entries when you ask" needs a ceiling on it.

There's a daily budget on how much of your journal the connector may read. A day of questions can only touch so much. So even a runaway assistant, a compromised session, or an unusually nosy afternoon can't quietly slurp your entire archive out through the question pipe. The budget resets tomorrow; your ten years of entries were never on the table today.

Why we're publishing this at all

Because vague privacy claims are how this industry launders behavior it doesn't want examined, and under-claiming is the house style here. We'd rather tell you the unglamorous specifics — spare key, transient decrypt, test suite, daily cap — and let you judge the machinery than hand you a slogan and ask for faith. One honesty note on the phrase itself: "burns after reading" is a promise about what's kept (nothing), not about what's seen — the AI does see an entry for the moment it answers you. That moment is the product. The burning is the point.

And if you read all of this and still think "I'd rather the machinery didn't exist at all" — good news, that's a setting. Private Vault deletes the spare key, the AI goes quiet, and we become cryptographically unable to read anything. The strongest argument that our claims are honest is that we hand you a free switch that takes our access away entirely.

An AI that remembers your journal is only worth having if you know precisely what it forgets. Now you do: everything, every time, by fire.