Goal — The knowledge layer AI agents work out of. An agent without durable memory starts over every time, drifts away from what was long since decided, and produces work that has to be done again. Infocluster is where that knowledge sits and stays: encrypted, separated per mandate, readable and writable by several agents at once — and visible, editable and provable to the human at any time. The benefit is that yesterday’s work still holds tomorrow.

Challenge — The hard problem of a shared knowledge layer is not storage. It is contradiction. As soon as several agents read the same body of knowledge — increasingly agents from different vendors — a contradiction inside it is worse than a gap: each runtime resolves it differently, one refuses where another proceeds, and neither asks. So what is sold is exactly what is missing then: one resolution and a trail — not the private heuristics of N runtimes.

Current state — Control plane, cryptographic envelope, search, storage and interface are built and run together as a whole — so far against fixtures; connecting real storage is the next step. 3 properties carry the whole thing:

What an instance cannot reach, it reports as unevaluated — never as green. Every access is carried as a receipt and is exportable, and the limit of that proof stands beside it: it evidences what crossed the interface, not what a foreign runtime did with it afterwards. And the strictest rule is built without a shortcut: where professional secrecy protects the client rather than the consultant, there is no acknowledge-and-proceed click — the protected party is not at the table, and the data model has no column for that click.

The human is not a spectator in this: he writes into it himself, changes and adds, whenever he wants. You land on a list, and when you want to see how things hang together you switch to the graph — which answers exactly one question per view, what is currently open or which agent reaches where, rather than showing everything at once and therefore nothing. in progress

What part AI plays — The leading one. The product is built for agents, not for humans with agents alongside: they tend and stock the body of knowledge — that is their work, not his. They reach it over a single, vendor-independent interface and under deny-by-default; access is granted per mandate and trust tier, and no agent is given the key to cross-mandate search. The human grants once, when access is given, not at each individual action; his supervision consists in always seeing what has happened and being able to withdraw access again. That is the deliberate exception to how the studio otherwise works, named rather than talked away: it shifts the weight from approving to proving.