remanence
Try troth
Remanence : troth

troth

A partner that's yours. Whatever model it runs on.

troth knows you, acts for you with its own hands, and keeps its word. What you settle with it on Monday is still there on Thursday, in any engine, without pasting a thing back in. Its memory, identity and rules live in one file on your machine.

Dexiosis, the handshake of a pledge. Troth is the old English word for it: to plight your troth is to give your word and keep it. Antiochus I and Herakles, Arsameia, 1st c. BC. Photo Ste.caneva, CC BY-SA 4.0.

What it doesthree things, together
01

It knows you.

Identity, memory, goals and the rules you set live in one local file. It reads the Claude Code and Codex history already on your machine, so day one is not day zero.

02

It acts for you.

A browser of its own, your tools, and a vault it can use without ever reading what is inside. Every command runs inside walls shaped to where it stands, and anything it touches can be put back.

03

It keeps its word.

What you told it never to do is checked before any model is asked. Swap Claude for ChatGPT, Kimi or a model on your own machine; the partner, and its promises, stay the same.

Monday → Thursday

Different model. Same partner.

On Monday you settle something on your Claude plan. On Thursday the quota is gone, you open a new terminal on another company's model, with no project open. It still knows. Nothing was pasted back in; the answer came out of a file on your own disk.

Monday · claude-opus-5
Thursday · kimi-k3
Three ways introth.one
The app

troth for macOS

Everything bundled, signed and updating itself. Voice, dictation into any app, secrets in the Keychain.

€229 once, free for seven days. No subscription.

Your engine

No second AI bill

It runs on the Claude, ChatGPT or Kimi plan you already pay for, any API key, or a local model with nothing leaving the machine.

There is no troth AI bill. Ever.

Your editor

The same mind, everywhere

Claude Code, Cursor, Windsurf, Cline and Claude Desktop can mount the same partner, one button each from the app.

Measured, not promised

83 of 100

Conversational recall on a LongMemEval-S slice, with the whole stack on one machine and no cloud anywhere in the loop. A cloud model reading the same memory scored 84: the memory, not the model, sets the score.