# Jev Recall

A memory filter that judges each candidate memory against a query with one calibrated yes-or-no question, so an AI assistant retrieves memories that are actually relevant even when they are not phrased like the request.

Jev Recall addresses the gap between keyword search, which only works when a memory happens to share words with the query, and semantic search, which only works when a relevant memory happens to read similarly to the query. It uses TypeSafe's Jev model as a classifier: every candidate memory is tagged with a short id inside one shared request, each gets a small yes-or-no question about whether it applies, and Jev returns a calibrated probability for each one in a single call.

Because the probabilities are calibrated, the tool keeps everything above a threshold rather than guessing at a fixed top-k, and returns nothing when no memory applies. Large memory stores are split across concurrent requests; an async variant, an inline mode that gives each memory its own question, and a shortlist mode that ranks everything with one multiple-choice question before verifying the top candidates are also provided. A benchmark included in the repository measures retrieval against keyword, dense and reranked baselines on a set of memories deliberately phrased unlike the queries that should retrieve them.

Website: https://github.com/samdotmak/jev-recall