Tenure

Most AI memory systems let context accumulate and rely on similarity search to surface what is relevant — which means the model never knows what it does not know, and stale decisions keep competing with current ones. Tenure replaces that with structured state: every belief carries an origin, a scope, a version, and a full history. When a decision changes, the old belief is retired and flagged so it is never suggested again, but the record remains for audit. Deployment is a single Helm chart or a one-line bash install that binds to port 5757 and presents an OpenAI-compatible `/v1` endpoint — no code changes required in VS Code, VSCodium, Open WebUI, or OpenClaw.

The differentiating feature is hard scope isolation. Probabilistic filters let context bleed across sessions when similarity scores drift close; Tenure uses structural boundaries so Project A’s memory is physically separated from Project B’s. The vendor pairs this with OIDC authentication, SCIM deprovisioning, and a real-time audit trail — every retrieval is logged with identity, timestamp, and the exact query that triggered it, recorded as it happens rather than reconstructed afterward. This architecture is specifically built for EU AI Act traceability requirements and multi-team governance scenarios where a shared vector store would be a compliance liability.

Tenure fits engineering teams running AI-assisted PR reviews who cannot tolerate a reviewer that has internalized last quarter’s abandoned design patterns, and for organizations where memory data cannot leave the building. The MIT license and no-call-home telemetry policy mean the codebase is auditable and the data stays on your infrastructure. The gaps show up when your deployment environment is not Kubernetes-ready, when you need a managed cloud tier, or when your team does not have the capacity to own the operational burden of a self-hosted state service — none of those paths are documented on the vendor site.