JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Integration walkthrough

Rerank agent memories with batched yes-or-no judgments

Jev Recall demonstrates retrieval by batching one relevance question per candidate memory and retaining memories whose calibrated yes probability clears a threshold.

Original by samdotmakRetrievalGitHub README and core moduleSource reviewed

Before you dive in

What you’ll find in the original

  1. Start with a query and candidate memories, assigning an id to each memory so answers can be mapped back without generated text.
  2. Batch one yes-or-no relevance question per memory in a single request rather than making a separate LLM call for every item.
  3. Choose and evaluate an inclusion threshold; probability-ranked decisions do not remove the need to test missed and irrelevant memories.
Worth knowing

Requires TypeSafe access for live runs. The README's latency, price, and example-retrieval comparisons are author-reported snapshots, not independent benchmarks or guarantees.

More guides →See the experiments →