When I worked as a doctor and researcher, there were AI tools we wanted to use but couldn’t give the sensitive information our work depended on. The difficult question was often quite specific: what would an assistant need to see to help with this task?
Vivendur grew from that question. We’re building a memory layer that an organization can hold separately from the AI tools it uses, with permissions for each tool and task.
Useful context, with a clear boundary
A study team preparing a recruitment update may need two facts: how many places are filled and when the next review happens. It does not need to disclose participant notes to answer those questions. A scheduling assistant joining later may need a different selection.
We want those decisions to be explicit, repeatable and inspectable. Someone responsible for the workflow should be able to understand what was shared, with whom, for which purpose, and under which rule.
What we mean by memory sovereignty
We use memory sovereignty to describe authority over the memory an AI workflow uses: where it lives, how it is shared, what is inferred from it, and when it stops being used. It is a design goal, not a certification or a category we claim to have completed.
We organize our work around five requirements:
- Custody. Your organization holds the main memory store, separate from the AI providers it chooses.
- Governed disclosure. Each context request is checked against policy, with a record of what was released.
- Inference governance. Inferences recorded by the memory layer carry their sources and can be reviewed, corrected or removed.
- Scoped erasure. Removal covers specified governed records and their tracked derivations, with a clear account of what remains outside that scope.
- Verifiable records. Receipts and ledger records can be checked, with the assumptions and limits of those checks documented.
These are the standards we want Vivendur to meet. Our status page distinguishes the prototype checks reported by our team from the production work still ahead.
Inferences need sources too
An assistant might turn a note about relocation into a guess that someone is changing jobs. That is an illustrative inference, not a fact the person supplied. If a memory system keeps such a guess, it should identify the source, retain the uncertainty, and give the responsible person a way to challenge it.
Our reference design treats stored inferences as records with source links and configurable disclosure rules. This cannot expose every conclusion held inside an external model. It governs the inferences that Vivendur records and tracks.
Ending access has limits
Revoking a tool’s permission would stop later requests through Vivendur. It would not recall information already sent to that tool. Provider conversations, caches, logs and exports remain subject to the recipient’s systems and agreements.
Within the governed store, we’re working toward erasure of selected records and tracked derivations. Production claims about cryptographic erasure will need evidence covering keys, backups, replicas and recovery paths. A certificate must state what it covers; it cannot, by itself, prove that every possible copy is gone. We describe those boundaries on our Architecture & verification page.
What we can show today
Our technical design and reference implementation put the core memory and policy rules into concrete form. Internal checks cover the behaviors summarized on our status page. The next phase is the production core, public verification tools and independent review.
The access challenge illustrates three decisions with fictional records and scripted responses. It is useful for explaining the idea; it does not enforce permissions in a real AI workflow.
What we are committing to
We will not monetize vault contents or build covert surveillance. We intend to remain independent of any single AI provider and to make the evidence for our security claims inspectable. Our commitments describe those choices and the obligations we want them to create.
We’re inviting organizations to discuss workflows and help shape the product before a pilot is agreed. The starting point is a conversation using fictional or redacted examples, not a request for sensitive records.