Coredence

See what your work says about you.

Coredence reads what you have already written, distils it into rules about how you actually work, and keeps every one of them pointed at the evidence it came from.

macOS 14 or later · Apple silicon · Notarised by Apple

What it produces

One rule, and the reason it exists

The engine reads what you have already written and writes down what keeps being true. Every rule carries the moments it was drawn from, so you can disagree with it about something specific.

This one turns up in seventeen of eighteen subjects. It is not about a project — it is about a person.

“Reaches for a formula before a fixed number, even when the number would do”

patternconfidence 0.69seen across 17 of 18 subjects

Drawn from

• “let’s not hardcode 40 here, make it scale with the level”
• “replace the magic numbers with a curve we can tune”
• “why is this a constant? it should come out of the formula”

Measured, not remembered

Rules come out of evidence, not out of you telling it about yourself.

Every claim has a source

Open any rule and read the messages it was drawn from.

Stays on your machine

One SQLite file in your home directory. No account, no server.

What it measures

Five kinds of claim, kept apart

A preference and a value are not the same thing, and a passing mood is neither. Collapsing them is how a memory feature ends up telling your assistant that you dislike a library you complained about once, on a bad afternoon, two years ago.

• Values

  What you will not trade away, however inconvenient. “Would rather cut the scope than lower the standard”

• Patterns

  What you do again and again, whether or not you would claim it. “Finishes the main path before touching the polish”

• Voice

  How you say things — length, directness, what you refuse to pad. “Says the uncomfortable version first and softens it after”

• Preferences

  What you reach for first when nothing forces your hand. “Prefers the boring dependency maintained for a decade”

• Current state

  What is true this month and will not be true next year. “Is deep in a launch and guarding the scope hard”

How it works

Download it, point it at your work, look

1. 1

   You point it at your work

   Conversations, commits, the corrections you made when something was wrong. It asks before it connects anything, and nothing is sent anywhere — all of it is read on your own machine.
2. 2

   It distils rules, and holds back what it cannot support

   Anything seen once is kept as a draft until it happens again. Two rules that contradict each other become a question for you, not a silent decision.
3. 3

   You look at the result

   The whole identity as one picture. Every rule opens onto the evidence it came from, and you can hide it, merge it, or argue with it.

Why you can trust it

What it refuses to do

The engine would rather say less about you than say something it cannot defend. Most of what follows is a decision to stay quiet.

• It will not keep a rule it has seen once. One sighting is a coincidence. It waits.
• It will not resolve a contradiction on its own. Two rules that disagree become a question you answer.
• It will not write a rule it cannot show you the reason for. Every rule keeps its evidence or it does not exist.
• It will not guess to fill a gap. A thin identity is honest. A padded one is not.
• It will not send anything anywhere. There is nothing to send it to.

Where it lives

One file, on your machine

The identity is a SQLite database in your home directory. There is no account, no server and nothing to sign in to. You can copy it, back it up, open it with any SQLite browser, or delete it.

The app that draws the picture opens that file read-only. It cannot change what it shows you, and that is a property of the file handle rather than a policy.

You never have to take that on faith. While it is running, ask the operating system what it has open: the answer is one database and no network connections at all.

$lsof -i -a -p $(pgrep -x Coredence)

$sqlite3 ~/.coredence/coredence.db '.tables'

The first command should print nothing. The second prints your whole identity, in a tool nobody here wrote.

Where this is going

What ships today, and what does not

The whole point of the product is that it only says what it can support. The same applies to the page you are reading.

1. 01

   Shipping now

   The engine on macOS and Windows, a window of its own on both, and the connection to the assistants you already use: read your work, distil it, look at it, hand it over.
2. 02

   Being built

   A view of how the identity changes over months rather than one snapshot.
3. 03

   Waiting to open

   The plan where we run the model is built into the engine; the sale is not open yet. The pricing page says what it will cost. Distilling today needs your own key.

Get it

One download, no terminal

The engine that reads your work ships inside the app. Nothing to clone, no Python to install, no command to run — open it and it asks what it may read.

Download for macOS

macOS 14 or later · Apple silicon · Notarised by Apple

• Open it It asks which sources it may read, and connects nothing until you agree.
• Leave it a couple of minutes The first pass reads a few hundred events. After that it keeps up on its own whether or not the window is open: reading every hour, distilling once enough has gathered.
• Look The whole identity as one picture, every rule opening onto the evidence behind it.

Windows works the same way: one download, a double-click, and both the window and your assistants know you. An identity is one file and moves between machines with you.

Questions

Before you install anything

Does anything leave my machine?

Only the passage being distilled, and only to the model you chose — with a model running on your own machine, not even that. There is no account and no server of ours: the app has nothing to talk to even if it wanted to.

What exactly does it read?

The sources you connect during setup, and only those — it asks first and connects nothing until you agree. While it runs you can check that it holds no network connections at all, with one command on the privacy page.

Can I disagree with a rule?

Yes, and that is the point of keeping the evidence. You can open a rule, read what produced it, and hide it, merge it, or answer the question the engine raised about it.

Does it work with my AI assistant?

Yes. Settings → Assistants lists the clients it finds on the machine — Claude, Claude Code, Cursor, Windsurf, VS Code, JetBrains IDEs, ChatGPT/Codex, Gemini CLI — and connects any of them with one button. The assistant queries the identity over MCP rather than being handed a file, and the connection is read-only.

Is it macOS only?

No. Windows has the engine and a window of its own, and an identity moves between the two machines as one file. The two windows are functionally equal: the constellation, the layers, search and evidence are in both.

What does it cost?

Using what exists today is free: you distil with your own key or with the Claude Code subscription you already have, and we are not in between. The only paid thing is the plan where we run the model for you when you have no account of your own — an addition rather than a wall around what you already have. That sale is not open yet; the pricing page carries the numbers.