Jev Dreaming Benchmark

Dreaming is a three-stage pipeline that mimics how biological memory works — filter, encode, consolidate. Incoming text chunks are first triaged by Jev, a probabilistic classifier that assigns a knowledge score to each chunk and drops irrelevant noise. Only fact-dense chunks advance to Gemini for memory extraction. Extracted memories are then judged by Jev again — each one is decided to Append (new fact), Extend (link into the knowledge graph), or Supersede (replace a stale belief). This benchmark measures how that hybrid Jev + Gemini pipeline compares to raw Gemini on speed, token cost, and memory quality.