Mneme HQ + NVIDIA Sentry
NVIDIA Sentry is NVIDIA’s security layer for AI agents. It runs on a network processor in the same server, in a hardware-isolated domain separate from the software the agent runs on, and monitors and controls the agent’s traffic. It is not Sentry.io, the error-monitoring service. Mneme is an open-source tool that stores a team’s architectural decisions in the repository and checks code changes against them, either inside a coding agent before it writes a file or in CI on a pull request. Whether the two can connect is a research question. No adapter exists.
Status: research
NVIDIA has not released Sentry yet. In October 2026 it said Sentry has “planned availability early next year”. Until it ships, and until NVIDIA documents how other software can feed it policy or read its events, there is nothing for Mneme to connect to. We have not had access to Sentry and have run no tests. Mneme has no partnership with NVIDIA.
What NVIDIA Sentry is
NVIDIA describes Sentry in a technical blog post and on its product page. It runs on BlueField-4, a data processing unit (DPU): a network card with its own processors. NVIDIA describes it as isolated from, and independent of, the host where the agent runs. It is built from NVIDIA’s DOCA (Data Center Infrastructure-on-a-Chip Architecture) software libraries:
- DOCA Flow: a per-agent network firewall that blocks connections to unauthorized destinations.
- DOCA Tower: a proxy built on Envoy, an open-source network proxy, that inspects allowed traffic at the application level. It uses Open Policy Agent (OPA), an open-source policy engine, to extend network access policies.
- DOCA Argus: inspects the host’s memory from outside it, without software on the host.
The blog post says Sentry keeps attested telemetry and tamper-evident records on the BlueField-4: activity logs that can be shown not to have been altered. Sentry is an optional part of NVIDIA’s OpenShell Reference Architecture, a design for running agents in sandboxes; the OpenShell runtime works without it.
How it differs from Mneme
Sentry sees network traffic and host memory. Mneme checks the content of code changes against recorded decisions. A decision such as “request handlers use the storage adapter” lives in source files, and Sentry does not read source files. A decision such as “agents call only the approved model endpoint” is about traffic, which Sentry does see.
| Question | What answers it |
|---|---|
| May this agent open this connection or send this request? | NVIDIA Sentry, where deployed, through DOCA Flow and DOCA Tower |
| Is the host or agent behaving like it has been compromised? | NVIDIA Sentry, through DOCA Argus and its telemetry |
| Does this code change respect a decision the team made? | Mneme, in CI or through a hook in a coding agent Mneme supports |
What we would investigate
NVIDIA says DOCA Tower can be extended with Envoy plugins and plans to publish core plugins; no dedicated public policy or plugin API is documented today.
- Write a policy from a decision. For a decision about which services an agent may call, a person writes the matching OPA policy for DOCA Tower. Sentry enforces the policy; Mneme keeps the record of the decision behind it.
- Link events to decisions. Connect a Sentry security event to the Mneme decision, and the decision’s version, that it relates to. This works only if Sentry’s events carry fields that make the link reliable.
- Report verdicts separately. Show Sentry security blocks and Mneme decision checks as different results.
What this does not cover
- OPA work does not carry over. Sentry uses OPA, but testing Mneme with OPA on its own would prove nothing about Sentry. See Open Policy Agent, also research.
- No source code checks in Sentry. Code-content decisions stay with Mneme’s agent hooks and CI.
- No infrastructure enforcement. Mneme does not enforce infrastructure or network policy.
- No certification or partnership. NVIDIA has not reviewed or endorsed Mneme.
FAQ
Is this about Sentry.io?
Can I use Mneme with NVIDIA Sentry today?
Do I need NVIDIA Sentry to run OpenShell?
Related reading
- NVIDIA OpenShell Reference Architecture: agent access policy is not architecture governance.
- NVIDIA OpenShell: planned integration status.
- Runtime governance: controlling what agents do while they run, and where that control stops.
- How enforcement works: where Mneme checks changes today.
Use what ships today
Sentry work is research only. Mneme’s shipped checks run in supported coding agents and in CI.
Mneme also exposes its Decision Index through a standard MCP server for interoperable access from compatible tools.