cairn-jev-lab
Experimental memory admission lab: Jev weighs a proposed memory against its source and a small policy says save, skip or defer.
Overview
Cairn Jev Lab tests what an AI should remember. You give it a source passage and a proposed memory; Jev evaluates the evidence, and a small, inspectable policy recommends save, skip or defer. The aim is to test a memory policy before letting it decide what an agent keeps, since a memory can sound plausible while changing what the source said.
The lab includes editable cases, a reusable JavaScript entry point with no runtime dependencies on Node.js 22+, and reports that keep both correct judgments and mistakes, retaining Jev's choices, probabilities and confidence alongside the policy's reason. A public recorded-only lab lets you browse results without a key. A paired developer-preview study on 100 synthetic cases reports that at threshold 0.40 Jev saved 39 of 50 intended memories versus 41 for Luna, with zero false saves for both, while noting labels were AI-authored without human review.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Does the source support this proposed memory? (evidence judgment)(inferred) | — |
| Choice | Policy outcome | save · skip · defer |
How it works
- Provide a source passage and a proposed memory
- Jev evaluates the evidence
- Deterministic policy maps results to save, skip or defer
- Report records probabilities and reasons, including mistakes
Strengths
- Inspectable policy separate from the model
- Keeps mistakes in reports
- Public demo needs no key
Caveats
- Developer preview
- Study labels were AI-authored without human review
Getting started
git clone https://github.com/Cairn-ink/cairn-jev-lab.git, set TYPESAFE_API_KEY in .env, then run the four-case sample.At a glance
- Type
- Service
- Status
- prototype
- Language
- JavaScript
- Runtime
- Node.js 22+
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Use it to test a memory policy before letting it decide what an agent keeps. The lab includes editable cases, a reusable JavaScript entry point, and reports that retain both successful judgments and mistakes. Node.js 22+, no runtime dependencies.
logicrw/awesome-jev-projectsTest what your AI should remember. An experimental, source-aware memory admission evaluator powered by Jev, with editable cases and inspectable results.
hellogumbo/awesome-jev
Frequently asked questions
What does it decide?
Whether a proposed memory should be saved, skipped or deferred, based on Jev's evidence judgment.
Can I try it without a key?
Yes, the public lab replays recorded results.
How did Jev compare to Luna?
39 versus 41 of 50 intended saves at threshold 0.40, both with zero false saves.
Related projects
Invalidation layer for agent memory: each fact gets a lease and Jev yes/no votes decide whether new evidence ends it.
ProjectbetaPythonEvidence-first experiment-memory skill for coding agents, with an optional Jev audit layer that checks provenance and claim support.
Agent skillbetaPythonUnix-style tools where Jev learns a repo's decision norms and judges each node of a code lead as a bug, lead or clean.
ProjectbetaJavaScriptOrca Lab plugin where Jev judges each agent command, passing routine ones, refusing dangerous ones and giving reasons for risky ones.
ProjectbetaTypeScriptnpm library where Jev judges whether an agent's recent steps are stuck on the same assumption; code returns CONTINUE, WARN, REPLAN or HALT.
ProjectbetaTypeScriptNode tool that crosschecks a folder of Markdown notes for contradictions and repeats, and ranks notes by any question, using Jev.
ProjectbetaJavaScript