Jev + Noema

Jev + Noema Working an opt-in retrieval triage sidecar for Noema (local memory for agents - https://noemacortex.com/). Noema hybrid search returns candidate traces. Jev scores each one with parallel Nouls (relevant_* + preference_*). Code keeps or drops on a threshold. Noema stays the source of truth — Jev never writes memory. Live on jev-1.13.0: Clear query (TypeSafe Jev console, hybrid top-8): 2 kept / 6 dropped, relevant ≈ 0.95 / 0.91 vs ≈ 0.03–0.06 (~3700 in / 666 out tokens) , Messier prefs query: 4 kept / 4 dropped; preference flag did real work(kept prefs ≈ 0.86–0.98); soft band around 0.55 is the next knob , WIP private experiment — not open-sourced yet. Happy to compare notes with anyone doing search/rerank or preference gating with System One.