How it works

You don't need any of this to use AIngram. But knowing the shape of it makes the rest of the guide easier to follow, and it shows why you can trust what's in there.

A notebook that only grows

AIngram is a notebook your AI writes in. Each note, called an entry, is short: what it's about (its subject, like people/ada), what sort of note it is (a fact, a decision, a correction), and a line or two of text.

Entries are only ever added. When something changes, your AI writes a new note that replaces the old one, and the old one stays in the record. So you can always see what your AI believed, and when.

Every entry also records:

• Who wrote it. Each AI that connects gets its own voice. Claude in your chats is the brain; Claude Code is Code. A voice comes from the AI's own key, so one can't pretend to be the other.
• When.
• Whether outside text was involved. If your AI had been reading a web page when it wrote the note, the note says so. That's a useful flag, because web pages can contain text meant to trick an AI.

Reading it back

At the start of a chat, your AI calls boot. That hands it three things:

1. The core: a few pinned notes it should always know, like your name and how you like to work.
2. The recent tail: its latest notes, so it can pick up where it left off.
3. A digest: one line on what the other voice wrote, such as Claude Code noting it shipped something.

During the chat it can recall more: everything on a subject, or a search by words.

Old notes fade, like real memory

Notes your AI never uses again slowly fade. A faded note drops out of the recent tail and sinks lower in searches, so it doesn't crowd out what matters now. It isn't deleted: it can still be found and read.

Using a note brings it back to full strength and makes it fade more slowly from then on. Pinned notes never fade. How fast notes fade is part of your policy.

Forgetting is real

When you tell AIngram to forget an entry, its words are removed from the database itself, not just hidden. A small marker stays behind saying "something was forgotten here", never what. Your AI is told that something was forgotten, so it isn't confused by the gap. Within a day the words are gone from your backups too.

Sealed, so tampering shows

Each entry is sealed with a fingerprint that includes the one before it, like a chain. Change any old entry, even by one letter, and every seal after it stops matching. The admin page can check the whole chain in a second, and your nightly backup checks it from outside the server, where nobody on the server can reach it.

Your wallet holds the rules

Who may write what, and how much, is set by your policy. Only your wallet can approve a policy, create a key for an AI, or forget an entry. Signing in to the admin page only lets you look; every change asks your wallet again, and shows you exactly what you're approving.