Vivendur · The sovereign memory layer

Bring AI into
sensitive work.

Put your organization’s knowledge to work with AI, with clear boundaries around what each tool should receive. Our approach combines private memory held by your organization with access defined for each tool and task.

Your encrypted vault
VivendurAccess you control
Selected contextYour AI tools

As featured in

USA News on Azin Etemadi and Vivendur rethinking who controls AI memory.USA News · Contributor feature · October 31, 2026

Read the feature ↗

Our approach

Start with what the task needs.

For example, a study team might ask an assistant to prepare a recruitment update. The number of filled places and the review date help it do that. Participant identifiers and appointment notes serve a different purpose. The same question comes up in a law firm’s matter files, a clinic’s records, and a public agency’s casework.

We’re building a memory layer between stored knowledge and AI tools, with access chosen separately for each tool and task. Sharing context with one assistant should not give every other tool the same access.

Our design brings those decisions into one place: check the request, select the useful context, and record what was shared. A different tool starts with its own permissions. Explore the architecture.

Try the three-round access challenge

How it’s designed to work

One gate between your knowledge and every AI tool.

The architecture centers on one policy gate. Each record can be shared, redacted or withheld for the task. Disclosure receipts are part of the design, recording what went out and under which rule.

How Vivendur is designed to work An encrypted vault held by your organization contains five fictional records, one of them an AI inference with confidence 0.58. Every record passes through one policy gate: recruitment figures and the review date are allowed, the coordinator’s name is redacted to “study lead”, participant notes are withheld, and the low-confidence inference is excluded silently because it falls below a 0.65 floor. Only the three approved items cross the trust boundary to the AI provider. The disclosure produces a signed receipt, entry 1042, recording three shared and one withheld, chained by hash to earlier entries including an illustrative erasure certificate for governed vault records and tracked derivations, not external provider copies. Responses return as untrusted input, and Vivendur’s own access is designed to be recorded in the same ledger. YOUR ORGANIZATION AI PROVIDER TRUST BOUNDARY Encrypted vault Held by your organization Recruitment · 24 of 40 Review date · Friday Coordinator · Dr. Reyes Participant notes Inference · target at risk · 0.58 Policy gate One path out. Every record checked. ALLOW needed ALLOW needed REDACT name → study lead WITHHOLD not needed EXCLUDE below 0.65 floor, silently AI provider Receives only the approved context Recruitment · 24 of 40 Review date · Friday Coordinator · study lead ↩ Responses return as untrusted input, checked before anything is remembered. signed receipt TAMPER-EVIDENT LEDGER #1040 · Disclosure To Assistant B · 1 shared Rule: reminders v2 prev 9c41… · signed #1041 · Erasure certificate Archived interview notes and tracked derivations in vault · design target prev 3e07… · signed #1042 · Disclosure receipt To AI provider 3 shared · 1 withheld Rule: study-updates v4 prev b51d… · signed VIVENDUR’S OWN ACCESS Designed to be break-glass only, dual-authorized, and recorded in the same ledger you can inspect. How Vivendur is designed to work An encrypted vault held by your organization contains five fictional records, one of them an AI inference with confidence 0.58. Every record passes through one policy gate: recruitment figures and the review date are allowed, the coordinator’s name is redacted to “study lead”, participant notes are withheld, and the low-confidence inference is excluded silently because it falls below a 0.65 floor. Only the three approved items cross the trust boundary to the AI provider. The disclosure produces a signed receipt, entry 1042, recording three shared and one withheld, chained by hash to earlier entries including an illustrative erasure certificate for governed vault records and tracked derivations, not external provider copies. Responses return as untrusted input, and Vivendur’s own access is designed to be recorded in the same ledger. YOUR ORGANIZATION Encrypted vault Held by your organization Recruitment · 24 of 40 Review date · Friday Coordinator · Dr. Reyes Participant notes Inference · target at risk · 0.58 Policy gate One path out. Every record checked. ALLOW Recruitment · needed ALLOW Review date · needed REDACT Coordinator → study lead WITHHOLD Participant notes · not needed EXCLUDE Inference · below 0.65, silently TRUST BOUNDARY AI PROVIDER AI provider Receives only the approved context Recruitment · 24 of 40 Review date · Friday Coordinator · study lead ↩ Responses return as untrusted input, checked before anything is remembered. TAMPER-EVIDENT LEDGER #1040 · Disclosure To Assistant B · 1 shared · prev 9c41… #1041 · Erasure certificate Vault notes + tracked derivations · design target #1042 · Disclosure receipt To AI provider · 3 shared · 1 withheld Rule: study-updates v4 · prev b51d… · signed VIVENDUR’S OWN ACCESS Designed to be break-glass only and recorded here.
Design illustration using fictional records. Follow the context from the vault through policy to the receiving tool. Architecture, verification and erasure scope · Try the access challenge.
Azin Etemadi, founder of Vivendur
Azin Etemadi
Founder, Vivendur

From our founder

When I worked as a doctor and researcher, there were AI capabilities we wanted to use but couldn’t, because we needed to protect the sensitive data our work depended on.

That experience led me to Vivendur. I want to make it easier to use AI in this kind of work, with a clear say over which information each tool receives.

Meet our founder

A closer look at access

Our founding essay, and three short notes accompanying the access challenge. All writing

Help shape what we build.

Bring us a workflow where sensitive information limits your use of AI. Together, we can identify the context a tool needs, the boundaries that matter, and a useful scope for a design partnership.

Discuss a design partnership