how it works

innernet is three things working as one: a memory at innernet.live, netti, who keeps it, and a door every AI tool walks through. Here is the loop, from a sentence in a chat to an answer in a different tool a week later.

1. something worth keeping is said#

You are in Claude, Cursor, ChatGPT or a terminal. A decision gets made, a preference comes up, a commit lands. The connected AI notices, guided by the project's capture protocol, and calls one of:

innernet_capture for the project ("we're going usage-based on pricing")
innernet_self_capture for you ("prefers short answers, works in TypeScript")
nothing at all, when it isn't worth keeping. Most things aren't.

In a repo, the git hooks do this for you on every commit and pull. See repo memory.

A capture is fast and dumb on purpose: it lands in the commit log in milliseconds, raw, before anything interprets it. A routing failure can never lose it.

2. netti folds it in#

On the server, netti's fold takes the captures since last time and consolidates them into the memory's dimensions: the pages that describe the project (vision, pricing, technical…). The fold replaces what was superseded rather than stacking it, and writes one dated line under ## history so nothing is lost. Personal captures fold into facts, each with a disclosure class and every source it came from.

Every fold is a commit. Your dashboard's inbox shows what each fold did (a new fact, a changed one, a line you said), and your memory's size in cells grows with it. See the inbox & the fold.

3. any tool asks#

A week later, a different tool, a fresh session. The AI calls innernet_context with your question in your words. innernet compiles a budgeted slice: the few sections that answer it (about 1,600 tokens at most, however big the memory), a few near-misses, the map by name, and the personal facts that bear on it, if any are shareable. See retrieval.

The answer builds on what was decided, and you never re-explained a thing.

4. you stay in charge#

See it. The dashboard shows every memory, every page, every fact about you, and where each came from.
Shape it. Tune the capture protocol, edit a page inline, move a fact's disclosure class.
Fork it. Branch a memory to try a direction, merge what wins.
Share it. One page as a link, a whole memory as a room, a skill to the community. Sharing
Take it back. Forget a page, a fact, or everything, in two steps.

what never happens#

No model runs on your machine. npx innernet wires config and hooks; every model call happens server-side, through one chokepoint.
No local copy. Nothing to sync, nothing to go stale. Every tool reads the same memory.
No private fact leaves without you. private and self facts stay inside innernet unless you allow an app to read them, and every time one leaves it is written down where you can see it.
No key in a chat. Clients sign in through your browser over OAuth 2.1. An AI asking you to paste a key is wrong; hand it the URL.