What Jev Means for AI Agent Memory

In September 2026, a new AI model went viral. It cannot chat. It cannot write code. It cannot write a single sentence. It does one thing: it makes judgments. Fast ones.

The model is called Jev, from a startup named TypeSafe AI. And if you run AI agents on a schedule, it matters to you even if you never touch it directly. Here is why.

What Jev actually is

Most AI models generate text one token at a time. Jev does not generate text at all. You send it a piece of text (or JSON) plus a set of typed questions, and it answers every question in one parallel pass.

There are three question types. A Choice picks one option from a list you define (up to 255 options). A Score places the input on a scale you describe. A Noul returns a calibrated yes-or-no probability. Every answer comes back with a confidence number attached. The full API shape is documented in the community reference.

The vendor's tagline is "a function call with frontier intelligence": unstructured state in, typed decisions out.

TypeSafe calls this a "System One" model, borrowed from Daniel Kahneman's Thinking, Fast and Slow. System Two is slow, deliberate reasoning, which is what reasoning LLMs do. System One is fast gut judgment. Jev is built to be the fast half. It was trained with a method they call RLCD, Reinforcement Learning for Calibrated Decisions, which teaches the model to be honest about how sure it is. In independent tests, when Jev said it was 90% sure or more, it was right 94 to 96 percent of the time.

The headline numbers: 70 to 500 milliseconds per call, $0.042 per million input tokens, and output tokens are free. That is roughly 100 to 500 times cheaper than frontier LLMs on decision tasks. The side-by-side comparison lays out exactly where each model type wins.

Why everyone is talking about it

The Wall Street Journal covered it this week: founder Diogo Almeida is a former OpenAI staff member who worked on ChatGPT before leaving in 2024. TypeSafe raised $40 million from DCVC and is reportedly in talks to raise $1 billion or more at a valuation above $10 billion. Almeida claims a quarter of the Fortune 500 already use Jev, at around a trillion tokens a day.

The technical crowd piled on for a different reason. Sebastian Raschka called Jev the "ChatGPT moment for classification": a general-purpose classifier that works on any text task without fine-tuning a custom model for each one. Within 48 hours of launch, developers had shipped a coding-agent guardrail, an MCP server, and interface clones. LangChain started testing Jev as an agent evaluator. Open clones like Laya appeared within days.

What it is actually good at

Strip away the launch noise and Jev is the component most agent systems build badly today: a fast, cheap, calibrated decision layer. Real uses already in production, collected here:

• Routing support tickets and messages to the right team
• Content moderation and spam detection
• Reranking search results (one legal benchmark went from 5% to 18% top-1 accuracy)
• Guardrails for coding agents (a tool called pi-warden blocked a destructive database-reset command 42 times across 17,000 judgments)
• Checking an agent's "done" claim against its transcript before trusting it
• Compacting context by scoring tool calls and dropping stale ones instead of summarizing
• Mining agent traces to find runs worth turning into reusable skills

Notice the pattern. None of these need a paragraph. They need a label, a score, or a yes/no. That is the job Jev was built for.

The part nobody is saying: judgments evaporate

Here is the gap. Jev makes a brilliant judgment in 200 milliseconds, and then it forgets it. Every call is stateless. An agent that judges a ticket as "billing, urgent, 94% confident" on Monday will pay to make the exact same judgment on Tuesday, because nothing was remembered.

This is the tax every automation operator already pays. Scheduled agents re-read, re-judge, and re-decide the same things every run. The judgment is fast now. The forgetting is still expensive.

The fix is to store the judgment where every agent can read it. When Jev (or any model) makes a call that matters, save the state, the question, the answer, and the confidence to shared memory. Next run, the agent checks memory first and only pays for a fresh judgment when the situation is actually new.

That is what Vilix AI is built for: one shared memory layer across every AI tool you use. Plan in Claude, build in Codex, and the context, decisions, and lessons come with you over MCP. It stores full conversation history, not just facts, so the real reasoning behind a past judgment is there when you need it. When two tools save conflicting info, the latest write wins, so a correction in one place becomes the truth everywhere. Your data is portable: export everything or delete it anytime.

The honest limits

Jev's "can't hallucinate" claim means it cannot return a value outside your schema. It can still pick the wrong valid option. Independent tests found it was the weakest of six models at rating answer quality, so it works best as a first-pass judge and reranker, not an oracle. It is also in early access: direct signups paused in September under demand, though it is served with no waitlist through Vercel AI Gateway and OpenRouter.

The takeaway for operators: the agent stack is splitting into layers. Fast judgment (Jev), slow reasoning (your LLM), and memory (Vilix AI). The teams that wire all three together stop paying the re-judgment tax. The ones that don't will keep making the same brilliant decisions, expensively, forever.

Get Started for Free

Free forever, no credit card. Paid plans start at $10/mo, with a 7-day Pro trial that needs no credit card.