Video unavailable. Open the source below.
NewAstra
An agent runtime optionally uses Jev to select relevant memories and detect explicit dismissal of saved lessons.
Added to Jevfast
How it uses Jev
Astra can use a configured judgment model, including Jev, to judge each candidate memory against the current user message. The runtime has separate Noul criteria for retrieval relevance and explicit dismissal. It applies the resulting judgments with configured thresholds, while preserving lexical fallback when model judgment is unavailable.
What Jev decides
Illustrative memory relevance and explicit dismissal judgmentsExample answers · not a recorded Jev response · Source ↗
Question 1 · 0
YOUR APP
INSTRUCTION
Does memory at state.candidates["0"] contribute a requested fact or applicable constraint to user_message? Apply state.policy.
STATE
Policy, truncated current user message, and a map of indexed truncated candidate memories. One Noul question is sent per candidate.
JEV · NOUL
YesNo
App workflow
- Collect candidates
The memory hook identifies candidate memories and builds a bounded request with policy, the current message, and indexed candidates.
- Judge each candidate
The configured judgment provider evaluates one Noul question per candidate for relevance or explicit dismissal.
- Select and report
Astra normalizes judgments, applies the matching threshold, records selected indices and falls back to lexical selection when judgment is unavailable.
Why it is interesting
It separates whether a memory applies to the current request from whether the user has invalidated the underlying lesson. A task-specific exception therefore need not erase a still-valid memory.