we are building the layer your thinking lives on — outside any one tool, kept by something that reads it, and yours to take with you.
every AI you use is brilliant for an hour and a stranger the next morning. you explain the project again. you explain why you rejected the obvious approach again. the reasoning was real, it was expensive, and it went nowhere — because there was nowhere for it to go.
innernet is the somewhere. one memory that belongs to you rather than to an application, reachable by every tool you already work in, and kept current by an agent called netti that reads what accumulates and writes down what it adds up to. you connect it once. after that the tools arrive already knowing.
almost everything in this category is memory for an application — installed by a developer so their product stops forgetting their users. that is a real problem and it is not this one. we started from the other end: the memory is keyed to a person, it lives outside every tool, and each tool is a guest that gets read access.
which is why we own no model and resell none. neutrality is easy to claim and hard to hold when your revenue depends on routing tokens through your own thing. ours doesn’t, and it never will.
most memory systems resolve toward the current answer: something new arrives, something old is overwritten, and the record of you changing your mind is quietly deleted. for a user profile that is correct. for a body of work it is the opposite — the approach you dropped in march is usually the one you need in july, when the reason you dropped it has changed.
so writes are commits, with a hash and a time and a message. contradictions do not get erased; they branch, and both stay alive until you merge them. what you used to think is as retrievable as what you think now.
encryption protects a memory from outsiders. it says nothing about what the model in front of you is allowed to know, and that is the boundary that actually matters the moment you connect an AI to your own life.
every fact carries how far it may travel, and the class is enforced when a tool reads. private and self-only facts never leave your own map, whatever is connected. it is a decision per fact, not a switch per account.
a memory has a survivorship requirement other software does not. if a notes app shuts down you lose a notes app; if a memory shuts down you lose the record, and the longer it worked the more you lose. “we won’t do that” is not an answer to it. the answers that count are structural.
a small team, working in the open about what is finished and what is not. we publish what shipped, we compare ourselves against the rest of the category in public — including the places others are better — and we write these letters out of the same memory the product keeps.