What does the memory wrapper change about a Jev call?

Only the state. ask_with_memory() takes the same state and questions as ask() and returns the same response, adding the customer's saved tickets under one key before the call and appending a new memory after the gate approves it.

How does the gate decide what is worth remembering?

A Noul question asks whether the request states a lasting fact about the customer, such as an edition, plan, or contract. At P(yes) of 0.5 or more the request and Jev's answer are saved. Facts scored 0.91-0.92 in the example; one-off details scored 0.07-0.10.

Why does the gate run after the answer instead of in the same request?

It judges the original request together with Jev's answer, and the page says that if your gate ignores the answer it could instead be a second question in the routing request via fan-out. As written, every call makes two sequential requests.

Can a saved answer reinforce a mistake?

Yes. The gate ignores facts that appear only in Jev's answer, so an answer cannot get a ticket saved by itself, but once saved its answer goes into every later call for that customer, mistaken inference included, and a later answer can repeat it.

How do I run the cookbook without spending API calls?

Set TYPESAFE_API_KEY for a live run, or download json_cache.json beside the script to replay the published numbers. Install with pip install ipython 'cooksafe>=0.2.0,<0.3.0'; the cached run needs no key.

What does the wrapper need in the state that plain Jev does not?

A customer ID. Plain Jev accepts a string as the state, but the wrapper reads state['customer']['id'] to find that customer's saved memories, so the example's ticket state keeps the customer object in the state.

What does the page suggest for a real deployment?

Replace the in-memory dict with a database keyed by customer ID, add a relevance question to pick which memories to include, ask whether a new memory replaces an older one, and map each route to a handler in code. The 0.5 threshold should be tuned on labelled examples.

OfficialJev Cookbooks

Simple Memory System Cookbook

Q: What does the wrapper need in the state that plain Jev does not?

A customer ID. Plain Jev accepts a string as the state, but the wrapper reads state['customer']['id'] to find that customer's saved memories, so the example's ticket state keeps the customer object in the state.

Q: What does the page suggest for a real deployment?

Replace the in-memory dict with a database keyed by customer ID, add a relevance question to pick which memories to include, ask whether a new memory replaces an older one, and map each route to a handler in code. The 0.5 threshold should be tuned on labelled examples.

A support-ticket cookbook that lets Jev decide which requests are worth remembering, then wraps each call so saved tickets ride along in the state. Six later messages route correctly with memory against two without.

See highlights

Open docs.typesafe.ai

Format: Docs
Published by: TypeSafe AI
Last verified: Oct 2026

Tagsofficial cookbook memory agents routing

Works forJev for Routing Jev for Classification Jev for Decision gating

Used inJev for Software & DevTools Jev for Customer Support

Languages: Python

Highlights

Two parts: a Noul gate that asks whether a request states a lasting customer fact, and a wrapper that adds saved tickets to later states.
The gate saves when P(yes) reaches 0.5; lasting facts scored 0.91-0.92 while airport and holiday small talk scored 0.07-0.10.
Six later messages reach the expected teams with memory against two without, because the same message routes differently per customer.
Each call makes two sequential requests, and the wrapper takes the same state and questions as ask() with the same return value.

Quickstart

python

def ask_with_memory(state: dict, questions: dict):
    saved = MEMORIES.get(customer_id(state), [])
    response = ask({**state, MEMORY_KEY: saved} if saved else state, questions)
    memory = as_memory(state, questions, response)
    if remember_probability(memory) >= REMEMBER_AT:
        MEMORIES.setdefault(customer_id(state), []).append(memory)
    return response

Watch out

The memory store is a plain Python dict that disappears when the process exits, and memory only grows: no size limit, relevance check, or expiry. The 0.5 threshold is a starting point the page says to evaluate against your own labelled examples. Numbers come from jev-1.13.0 on 2026-09-28.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Facebook: Intake Router keeps diagnosis out of the answer space at the front desk.

FacebookClinicnotes

YouTube: Jev explained for routing and judging, with a comparison against LLMs.

Hacker News: jevmem keeps Claude Code project memory by asking Jev which turns are worth saving.

avinashjetwani61 pointsJevmem – automatic project memory for Claude Code, built on JevRead the thread on Hacker News

Hacker News: Vercel on using Jev for defined decisions and GPT-6 Astra when the job also has to write or use tools.

flashbrew4 pointsTypeSafe AI Jev vs. GPT-6 AstraRead the thread on Hacker News

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Jev Handson

Korean lecture-lab code for a three-part Jev course plus a CCTV extra: first calls, agent-specific routers for Claude Code, Codex, and OpenClaw, keyless replays, Colab notebooks, and OneJev video judgment.

Jev Cookbooks#community#korean#education

Community

jkf87/jev-handsonJupyter Notebook

Twitch Chat Filter Cookbook

A live chat filter that asks three questions per message — category, toxicity, and spoilers — then turns the answers into show, highlight, dim, hide, or blur in plain code. The 29-message example hides seven, blurs two, and highlights six.

Jev CookbooksDocs#official#cookbook#moderation

Official

LAVOIR: Teaching a Single-Pass Decision Encoder When and What to Ask with Amortized Value of Information (arXiv 2609.30706)

LAVOIR extends the Laya single-pass decision encoder so one forward pass returns both the decision distribution and, for every candidate missing slot, the expected gain from asking the user about it. Its question policy matches a greedy oracle on seen schemas and is 14.1 points more accurate than never asking.

Jev Models & ReimplementationsArticle#community#research#arxiv

Community

moganai/lavoirPython

Jev-Style 2B Decision v3 GGUF

The v3 Jev-Style decision model at 2B in GGUF: F16 3.78 GB and Q8_0 2.01 GB match PyTorch FP32 on 100% and 99.7% of parity rows, 73.6% on JevBench v1.4.1 public items.

Jev Models & Reimplementations#community#open-models#local

Community

Cookbook: Self-Consistency for Nouls

A 14-Noul claims-triage rubric over one insurance claim, repeated 15 times: mean probability standard deviation of 0.0102 at 111 ms per call, with a 0.30 to 0.70 band for human review.

Jev CookbooksDocs#official#cookbook#consistency

Official

Claude x Jev

A free Claude Code skill that puts Jev in front of the expensive model: thirteen commands sort, check, score, gate and verify with Jev, so Claude reads only the unsure slice and writes the output. Needs python3 and an OpenRouter key.

Jev Agents & MCP#community#python#claude-code

Community

charlesdove977/claude-x-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Retrieval with Decisions: Filter, Rerank, and Verify

Oct 12, 2026

Retrieval is a sequence of bounded judgments. How a decision model slots into passage filtering, semantic find, reranking, and citation checks, and what stays in deterministic code.

Jev Cookbooks#retrieval#reranking#rag

Read article

One Request, Many Questions: Batching, Fan-Out, and Composite Scoring

Sep 12, 2026

Jev bills input tokens and gives output away, and questions in one request run in parallel. The design consequence is simple: ask more per call, then decide in code.

Jev Cookbooks#batching#patterns#search

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai