Jev judgments for governed agent memory
As of September 22, 2026, Heartwood Memory's Jev integration is in early access with a waitlist; it is not generally available. Heartwood Memory is adding an opt-in judge that uses Jev, TypeSafe AI's System One model, to rank recalled memories and to check AI-written memories against their sources before they're saved.
It is off by default and offered only to hosted Team and Professional organizations that turn it on in writing. Community, self-hosted and air-gapped deployments never call Jev, and Heartwood's local ranker stays ready as the fallback.
What the Jev judge will do
Rank recall results
The judge will score how well each policy-cleared memory answers the query, so the most relevant memories come first. Heartwood's local ranker runs alongside every judged recall, and latency-critical recall paths stay on the local ranker.
Check AI-written memories before they're saved
Before an AI-written memory is saved, the judge will check it against the source passages it cites. A memory is accepted only when Heartwood's built-in check and the judge both find it supported. A contradiction or an unsupported claim is rejected, and any other disagreement goes to review. The judge can make the check stricter, never looser.
What is Jev?
TypeSafe AI describes Jev as its first System One model. Instead of writing text, it answers structured questions (a yes or no, a choice among options, or a score) with a probability for each answer. See TypeSafe's announcement and documentation for details.
How it fits Heartwood's governed recall
This is how the early-access judge is designed to work. Step 1 ships today; the rest is the plan.
Policy filters first. Records the calling agent isn't cleared to use are removed before anything is ranked. This part ships today.
Your organization's switch, then the egress check. The judge runs only if your organization has turned it on, and only after Heartwood's egress check allows an outside model for these records.
All or nothing. If any candidate is above your egress ceiling or labeled as personal data, or the request matches a secret pattern, nothing is sent and the whole recall uses the local ranker.
The local ranker stands by. It runs alongside the judge. If Jev is slow or unavailable, recall returns in the local order.
The call is recorded.
explain_recallis designed to show which ranker ordered the results, the egress decision, the model and TypeSafe request ID, a hash of what was sent, and each memory's probability.
What leaves your environment, and when
Only for a hosted organization that has opted in, and only when every check above passes.
| What | Sent to a judge? | What Heartwood keeps |
|---|---|---|
| The recall query and the candidate memories being ranked (clipped), when every candidate is cleared | Yes, to TypeSafe AI (US). Vercel AI Gateway is a backup route only, used when TypeSafe is down or too slow, with the same request and the same checks | The explain record: ranker, egress decision, model, request ID, a hash of what was sent, one probability per memory |
| An AI-written memory and the source passages it cites | Yes, to the same two processors, on the same terms | Judge identity, label probabilities, egress decision and the final outcome |
| Records the calling agent isn't cleared to use | Never. They're removed before ranking | — |
| A recall where any candidate is above your egress ceiling or labeled as personal data | Nothing is sent. The whole recall uses the local ranker | The reason |
| A request that matches a secret pattern | Nothing is sent, to either processor | A refusal reason, never the matched value |
| Organizations that haven't opted in; Community, self-hosted and air-gapped deployments | Never | — |
| The judge's answers | — | Used for ranking and checking only. Never used to train or fine-tune any model |
Both processors are named here because both can receive your data once you opt in. Retention on their side follows their own terms: TypeSafe AI legal and Vercel AI Gateway. Heartwood doesn't describe this as a zero-data-retention arrangement.
What stays the same
- The self-hosted core makes no Jev calls, with or without this integration.
- Recall without the judge works as it does today: policy filters first, the local ranker orders the results, and explain_recall shows why.
- Nothing is required of you. If your organization never opts in, nothing changes.
What we'll publish, and when
We haven't published any accuracy, speed or cost figures for the Jev judge. We'll publish results only after measuring them on Heartwood's pre-registered recall test, with how each number was measured shown next to it. Results will be linked from this page and from our measurement receipts.
Join the early-access waitlist
For hosted Team and Professional organizations. Tell us which agent and which recall you'd put behind the judge. We target a response within one business day.
Questions
Does Heartwood Memory work with Jev?
As of September 22, 2026, Heartwood Memory's Jev integration is in early access with a waitlist; it is not generally available. Heartwood is adding an opt-in judge that uses Jev, TypeSafe AI's System One model, to rank recalled memories and to check AI-written memories against their sources.
What will the Jev judge do?
Two jobs. It will reorder the memories recall returns so the most relevant come first. It will also check whether a memory an AI wrote is supported by its sources before that memory is saved.
Is it on by default? Who can use it?
No, it's off by default. It's offered only to hosted Team and Professional organizations, and only after an organization turns it on in writing.
What data is sent to TypeSafe, and who processes it?
Nothing, unless your organization opts in. After that, the recall query and the candidate memories being ranked, or an AI-written memory and its cited sources, go to TypeSafe AI in the US. Vercel AI Gateway is a backup route only. Both are named processors. Nothing is sent if any record is above your egress ceiling, labeled as personal data, or matches a secret pattern. Retention on their side follows their own terms, and Heartwood doesn't describe this as zero data retention.
Can I use it with Community or self-hosted Heartwood?
No. Community, self-hosted and air-gapped deployments never call Jev. Recall there always uses Heartwood's local ranker.
What happens if Jev is slow or unavailable?
Recall still returns, ranked by Heartwood's local ranker, which is designed to run alongside every judged recall.
I can call Jev myself. Why use it through Heartwood?
You can. Jev is a public API, and other memory tools offer Jev ranking too. Heartwood's plan is about the governance around the call: only memories your policy clears can be judged, the feature is opt-in per organization, each judged recall is recorded with the model that ranked it, and both processors are named.
Do I get Jev API access through Heartwood?
No. Heartwood uses Jev inside its own service. The integration doesn't give you access to Jev or TypeSafe's API.
How much does it improve recall?
We haven't published figures. We'll publish them only after measuring on our pre-registered recall test, with how each number was measured shown next to it.
Does Heartwood train models on Jev's answers?
No. Heartwood doesn't use Jev's outputs to train or fine-tune any model.
Is Heartwood Memory affiliated with TypeSafe AI? Is it connected to other Jev memory projects?
No to both. Heartwood Memory is made by Edukas Solutions LLC and is not affiliated with or endorsed by TypeSafe AI. Third-party Jev memory projects are not made or endorsed by Heartwood Memory; its own Jev integration is the one described on this page.
When will it be available, and what will it cost?
We haven't set a date or a price, and it isn't part of any current plan. Join the waitlist on this page, and we'll contact you when it opens for your plan.
Jev is a model from TypeSafe AI, Inc. Heartwood Memory is made by Edukas Solutions LLC and is not affiliated with or endorsed by TypeSafe AI.