Notes · from the person building Mneva
Memory you can read.
The strangest thing about AI memory, as it is mostly built today, is that you cannot read it.
Think about what a memory is between two people. If a friend remembers something about you, you can ask them what they remember. They will tell you, in words, and if they have it wrong you can correct them, and if it is something you would rather they let go of, you can ask, and a good friend will. The memory is legible, correctable, and closable. Those three properties are not features of friendship; they are what makes the remembering tolerable at all.
Most machine memory has none of them. The common pattern is a system that quietly decides which fragments of what you said are worth keeping, stores them in a form you never see, and reveals what it kept only by accident — a reply that knows something you do not remember telling it, phrased in words that are not yours. When that happens it rarely feels like being remembered. It feels like being filed.
The failure is not that the machine remembers. People want to be remembered — it is one of the oldest wants there is. The failure is that the remembering is passive and opaque: you did not choose it, you cannot inspect it, and you cannot end it, except by a settings page that offers you all-or-nothing.
Mneva is built on the opposite position, and the position has a name inside the product: Holding. What Mneva keeps, it keeps as threads you can open and read — the promise you made, the decision you postponed, the worry you named — in your own words, not in a machine's paraphrase of you. Each thread can be edited when you said it badly, closed when it is finished, and deleted when it should never have been kept. And when a thread matters again, it comes back to you at the moment it matters — not as a score, a streak, or a chore, because remembering is not a game and you do not owe your own memory a daily login.
There is a second half to the position, and it is the harder half: what is sensitive runs on consent. A system that notices how you are, not just what you asked, is handling the most protected category of information there is. Mneva asks for that consent explicitly, inside the conversation, and withdrawing it is as easy as giving it. The details, including the honest limits, are on the Privacy page — which was written to be read, not to be scrolled past.
Mneva is in early access and under active development; unfinished parts are labelled unfinished, on every page where they appear. But this position is not unfinished. It is the foundation, and it does not change: memory you can read, in your own words, built to stay yours.