Your AI memory. Sealed. Yours.
Every AI interaction, every agent decision, every policy verdict is archived in an open, signed, portable format on your own device. Exportable. Searchable. Verifiable by anyone with the public spec. This is the substrate for what comes next: Recall in mid-2026, your Personal Model in Q3 2026.
v1.0 published
On your device
Tamper-evident
Fortress and above
Providers forget. You remember.
AI providers are optimized for the 500 million-user average. They cannot know you. They will not carry your history across models, vendors, or companies. When a provider rotates its memory layer, your context is gone.
Vault is built for the opposite incentive. One user. One history. One copy, owned locally, portable forever. The structural differentiator against any frontier model is not intelligence, it is longitudinal context that cannot be reproduced at planetary scale.
Three stages. One continuous archive.
Vault is not a feature. It is the substrate. Every stage builds on the one before. You are not migrating data between products. You are accumulating a record that becomes more valuable every month.
Vault. The archive.
Recall. The search layer.
Personal model. The edge.
Four capabilities. Shipping now.
Vault v1.0 is live in every Vigil installation. The four things it does are deliberately small. The value is in doing them correctly, durably, and in an open format that outlives us.
Archive. Every exchange.
Seal. Each record signed.
Export. Any time.
Replay. Any incident.
Integrated with the Cloud Agent Registry.
Every agent that routes through Vigil Gateway receives a TAP-issued certificate. The Cloud Agent Registry is the directory of those agents. Vault now indexes every record by agent identity, not just by session. Your archive knows which agent did what, when it was authorized, and when its authority ended.
Agent-keyed memory, not session-keyed.
Traditional logs key by user and timestamp. That is enough for a single-model world. It falls apart the moment you have a calendar agent, a banking agent, a research agent, and a coding agent all operating on your behalf. Vault keys every record by the TAP agent identity that produced it.
The result is a per-agent history you can reason about. Every agent has a lifespan. Every agent has a scope. Every agent has a revocation event if something went wrong. Vault carries all of it.
- 01
Agent-scoped queries
"Show me everything my research agent did in the last 30 days." One filter, complete history, cryptographically sealed. Works identically whether the agent is active, dormant, or revoked.
- 02
Revocation-aware records
When an agent is revoked via VARP, Vault flags every record produced by that agent from the revocation moment forward. The history remains searchable. The authority state is clearly marked. No silent loss of context.
- 03
Cross-surface agent continuity
A single agent identity persists across desktop, cloud Gateway, and partner B2B2C deployments. Vault consolidates the full cross-surface history under one agent ID. No more piecing together logs from three vendors.
- 04
Registry-backed attribution
Every Vault record carries the signing authority, the issuing TAP certificate, and the revocation endpoint. The record tells you who vouched for the action at the moment it happened. Attribution is not reconstructed, it is recorded.
$ vigil vault query --agent agt_a47e9c1b \
--since "30d" --include-revoked
✓ agent_id agt_a47e9c1b
✓ registry vigil-authority/v2
✓ principal user@example.com
✓ scope chat.read, chat.write,
stripe.read
! status REVOKED 2026-04-18T09:12Z
✓ records 847 total
✓ chain_ok true (SHA-256 verified)
# Records in window:
# pre-revocation: 801 (VOAF sealed)
# post-revocation: 46 (flagged, quarantined)
$ vigil vault export --agent agt_a47e9c1b \
--format voaf-m --out agent-trail.jsonl
✓ Wrote 847 records · 12.3 MB · SHA-256 chain intactVOAF. Open by design.
The Vigil Open Audit Format is the substrate Vault writes to. One JSONL line per record. Hash-chained. Signed. Third-party verifiable. Published as a tagged specification so other tools can read and write it without Vigil.
Open spec, dominant issuing authority.
The format is open so adoption compounds. Any tool that handles AI audit evidence can read a Vault export. Any regulator can verify a record without asking us. Any partner can issue records in the same format without licensing.
The economics are open spec, proprietary issuance. Vigil is the dominant origin of VOAF records in the wild. The comparison is DigiCert for certificates or FICO for scores. The format is the rail. The issuance is the business.
github.com/vigilsec/voaf-spec carries the tagged v1.0 specification.
// VOAF v1.0 record · one JSONL line
{
"voaf_version": "1.0.0",
"record_id": "voaf_7f2a81c9",
"agent_id": "agt_a47e9c1b",
"issuer": "vigil-authority/v2",
"ts": "2026-04-22T08:14:32Z",
"surface": "gateway.openai",
"request": { /* parsed intent */ },
"response": { /* action scored */ },
"verdict": "allow",
"agency": 0.34,
"policy_rev": "rev_412",
"prev_hash": "0x3b7e...c9f1",
"self_hash": "0x4c8f...e2a1",
"signature": "0x9a1d...b3e7"
}Retention scales with the coverage envelope.
Every paid tier ships Vault. What changes is retention depth and access to the layers above it. Personal model access begins at Citadel as a preview, lands fully at Sovereign when it ships in Q3 2026.
Your AI life is a record. Start keeping it.
Vault runs from the day you install Vigil. Nothing to configure. Nothing to opt into. Every interaction sealed from the first one. The earlier you start, the deeper the substrate when Recall and the Personal Model arrive.