Automaton Nebula
The questions that come up most, answered with what the product does today. If something is not tested yet, it says so here.
A context vault per client. You keep a permanent profile of who you are and how you work, and on top of it one workspace per client or project, with dated facts. When you open a connected AI it receives the profile and the active workspace: eight to twenty current lines, not the whole history.
It sees your profile and the current facts of the workspace you have open. It does not see your other workspaces, expired facts, or the ones it proposed itself and you have not confirmed. For an undecided conflict it does see both values, marked as such and with an instruction not to assume either: that is the only way it asks you instead of choosing on its own. The panel shows what goes and what stays out, with the reason, before anything is sent.
No, unless you allow it from the panel. A connector starts without permission to delete, because forgetting takes the fact and all its versions with it and cannot be undone. It also cannot see workspaces you mark as sensitive.
Neither is injected. The AI is told there is something undecided, with both values in view, so it asks you instead of choosing on its own. When you decide, the winner stays current and the other stays in the history with its date.
With anything that speaks MCP, and five have now been checked one by one: Claude (with a connector and with a device token), ChatGPT, Grok, Vibe — formerly Le Chat — and Codex. Grok works on a standard account; Vibe on its free plan, though there you also have to switch it on inside the chat. In ChatGPT you first have to switch on developer mode in the settings, and then it works on Plus and on Pro; the free plan does not have it. Cursor, Claude Code and Windsurf take the same path as Codex, a device token, but have not been tested one by one. Perplexity allows a custom remote connector on its enterprise plan, and its local one is Mac only. And for any AI without a connector there is the button that copies the context block to paste as the first message.
On our own server in Germany, inside the European Union. You take the whole vault whenever you want, in one file with a documented, versioned format: workspaces, facts with their history, conflicts, archive and the decisions you already made. And it deletes at once, without writing to anyone.
No. Nothing here is trained, nothing is sold and no third party receives it. What does leave is what you send to the AI you connect: those lines travel to that AI’s servers, and you decide that workspace by workspace.
Nothing is deleted. What is already stored keeps being readable, searchable and exportable; only new writes wait, until you are back within the plan limit. The free plan does not expire: one complete client workspace, one hundred facts and two connectors.
Nebula