JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Research writeup

Jev changes a lot in memory & context engineering. Here's exactly how.

Supermemory’s founder tests Jev in reranking, semantic chunking, memory prefilters, and agent-memory hooks. Dedicated rerankers sometimes win on quality or cost, and filtering sentences can lose needed context.

Original by Dhravya ShahRetrievalX articleOriginal published Source reviewed

Before you dive in

What you’ll find in the original

  1. Jev improved BM25 reranking on the tested BEIR datasets, but a dedicated reranker could be cheaper or score higher.
  2. Sentence-boundary decisions produced promising chunks in internal multilingual tests, at higher cost than simpler chunking methods.
  3. Filtering sentences before memory extraction removed 58% of content tokens in an internal benchmark but risked losing context; deciding whether to retrieve memory in an agent hook is a separate use case.
Worth knowing

The benchmark comparisons and internal results are reported by the author, not independently reproduced by JevMade. The agent-hook discussion is exploratory, not a measured production outcome.

More guides →See the experiments →