Are the benchmark numbers real?

The repo ships bench data with 238 fictional memories and 18 labeled requests where the decisive memory is phrased nothing like the request, plus a harness that runs BM25, embedding, reranker, and LLM baselines. Pointer mode passes most requests and matches the Sonnet reference arm, but the comparison notes LLM costs above exclude prompt caching.

Why not just use embeddings or an LLM re-ranker?

Keyword search needs the right terms and embeddings need similar phrasing, so both miss cases like an almond-tart order versus a nut-allergy memory. An LLM re-ranker catches those but charges LLM latency on every turn; Jev makes the same include-or-not call as a classifier.

Do I need Anthropic and OpenAI keys just to use it?

No. The library itself needs only a TypeSafe key. The other two keys are only for running benchmark baselines and for recording the demo replays; the demo replays recorded runs by default and keeps entered keys in browser local storage.

What happens when no stored memory is relevant?

Because probabilities are calibrated, it keeps everything above the threshold instead of guessing a top-k, and result.texts can be empty. Every memory still appears with its probability in result.all_scores.

How do I install and call it?

uv sync installs the library and dev dependencies, and a TypeSafe key is all the library needs. In Python, Recall() reads TYPESAFE_API_KEY, and recall.filter(query, memories, threshold=0.5) returns result.texts, result.all_scores, result.cost_usd, and result.latency_s; an async afilter is also available.

What are the request modes, and how do they differ?

pointer mode puts every memory in one shared state tagged with a short id, each with a tiny yes/no question, which is the benchmark winner. inline puts each memory's text in its own question, and shortlist ranks everything with one multiple-choice question and then verifies the top candidates.

How does it handle a large memory store?

Large stores are split across requests and sent concurrently; afilter is the async version for that. Because each memory is judged independently with a calibrated probability, you keep everything above the threshold rather than picking a top-k, so nothing is returned when nothing applies.

CommunityJev Tools & CLIs

Jev Recall

Python memory filter that judges every stored memory against the current request with one calibrated yes/no per memory in a single Jev request, keeping everything above a threshold.

See highlights

Open github.com

Stars: 38
Language: TypeScript
License: MIT
Published by: Community
Author: samdotmak
Last verified: Oct 2026

samdotmak/jev-recall

User repository on GitHub

View on GitHub

Retrieve by relevance, not resemblance: filter an AI assistant's memories with TypeSafe's Jev

Forks: 3
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Jev Recall gives you LLM re-ranker quality retreival at semantic search prices and speed. Given a user query and a pile of memories, jev recalls returns the most relevant memories, judged by TypeSafe's Jev model, one calibrated yes/no per memory, in a single request. Try the live demo → · replays real runs, or runs live with your own API keys. With LLMs, writing memories is easy. Knowing which ones to retrieve is the hard part. Today you get two options that don't work because:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Moderation

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Memories share one request state tagged with short ids, each with a tiny yes/no question returning a calibrated probability.
Keeps everything above a threshold instead of top-k, so nothing is returned when nothing applies.
Pointer, inline, and shortlist modes with concurrent sharding for large stores, plus an async afilter.
Ships unit tests against a mocked Jev API, a bench harness with baselines, and a Next.js demo site.

Watch out

Judges every memory each turn, so per-turn cost and latency scale with store size; the one reported miss needed outside knowledge to connect. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jevegis

Jevegis wraps Jev into a single guardrail call for LLM apps: one request covering prompt injection, jailbreaks, credential and PII leaks, unauthorized actions, malicious code and indirect injection in retrieved documents, plus Trust & Safety moderation. It returns probabilities rather than booleans.