Every model,
fitted to you.

01

The signed ledger

If a system's memory of you is a file it keeps rewriting, you cannot tell what changed, who changed it, or rebuild it later.

Think of a bank statement. Your balance is worked out from a list of transactions that is never erased. A mistake is fixed with a new entry, not by rubbing out an old one.

Exo's ledger works the same way. Every event is added to the end: what you said, what you ruled, what agents did, and whether you accepted, edited or rejected it. Each entry carries a digital signature (Ed25519), a mark only its author's private key can make, so anyone can check who wrote it.

Each entry also holds a hash of the one before: a short fingerprint that changes completely if one character changes. Edit an old entry and every fingerprint after it stops matching.

Your profile, memory and skills are worked out from the ledger, like the balance. When a better model arrives, it reads the ledger from the start and catches up.

02

The personal store

What Exo knows about you should belong to you, move with you, and be quick to search.

So it lives in one folder you can copy: a one-page summary called the pack, an index called the router that says where each kind of knowledge lives, beliefs (claims about you, each with its evidence and how sure Exo is), imported items such as notes and bookmarks, and the ledger itself.

Search works two ways at once. Keyword search finds exact words. Embeddings find meaning: a small model turns each passage into a list of numbers, so passages about the same idea sit close together even when they share no words. "My flight home" finds a note about "the return trip".

Exo merges both results, follows links to related notes, and stops at a token budget. A token is a piece of a word, and the budget caps how much text the model is handed. The embedding model runs on your machine, so your text is never sent away to be searched.

04

Fresh context every turn

Most chat tools send the whole conversation to the model every time you write. It grows until early details get lost, and old mistakes keep coming back.

Picture joining a long project. You do not reread every word of every meeting. You read the minutes, then listen closely to the last few minutes of talk.

Exo does that. Each turn, one message from you and one reply, is a new call to the model. Exo builds it from a summary tree: small groups of messages are summarised, then groups of summaries are summarised, like chapters under a table of contents.

The latest messages go in word for word. The turn waits until the summaries are up to date, so nothing falls into a gap. Summaries are written in parallel, with live chat ahead of imported history.

05

Gates by reversibility and exposure

Ask for approval on every step and people stop reading and approve everything. Ask for none and one bad step can cost a lot.

A family can let anyone move the furniture but need two signatures to sell the house. What matters is not the size of the step. It is whether it can be undone, and who else it reaches.

A gate is a check that runs before an action, and Exo's gates ask those two questions.

Private steps that can be undone run at once and are logged. Steps that cannot be undone, or that are public, involve money or reach other people, wait for your answer, with a link and a suggested reply. Every outside action is tied to a named person. Rules that protect other people sit outside the model of you, so learning your taste can never argue Exo out of them.

06

Decision models for small choices

Many choices in a session are small: which tool, which model, which label, whether to open a panel. Asking a large model each time is slow and costly, and it answers in prose that then has to be read.

A firm does not ask a senior partner which folder a letter goes in.

Exo sends these choices to a decision model: a small model that is given a fixed list of options and returns how likely each one is to be right, such as "tool A 0.82, tool B 0.11". The numbers show how sure it is, so Exo can act when it is confident and ask when it is not.

The default is Laya, which runs on your machine, free and private. Users can choose TypeSafe Jev, Perplexity's decider or OpenAI's Decisions API instead. A decision model may make a gate stricter, never looser. Every decision is logged so the learning loop can score it.

07

A standalone core with a protocol

Most assistants are tied to one app. Move to a phone or a new screen and you start again.

Email solved this long ago. One mail server holds your inbox, and any mail app shows the same inbox, because they all speak the same standard.

Exo's core, exod, is a small Python program that does the real work. It runs each turn, picks which model should answer (the model router), calls tools and keeps sessions.

The terminal, web, desktop, chat channels and a future device are clients: screens that talk to the core through one protocol. Everything that happens is sent as an event with a fixed shape and a number in sequence. A client that drops off asks for every event since the last number it saw, so nothing is lost. A cap on lines of code, checked by tests, keeps the core small enough to fit on a device.