NewRikkaHub SillyTavern Android
An Android chat client optionally uses Jev to screen memories for relevance and expose a judge tool to its main model.
Added to Jevfast
NewHow it uses Jev
The Noul relevance probability is compared to 0.5; eligible memories are injected into the conversation. If screening is unavailable or returns no relevant memories, the transformer falls back to semantic or lexical retrieval.
What Jev decides
Illustrative source-defined Jev questionExample answers · not a recorded Jev response · Source ↗
Question 1 · m0
YOUR APP
INSTRUCTION
这段对话与下面这条记忆是否相关？相关指的是：这条记忆对理解当前对话、或对当前该说什么有帮助。身份类记忆（称呼、偏好、习惯）即使当前对话没直接提到也算相关。只是同一话题领域但没有实际帮助的，才算不相关。记忆内容：<memoryContent>
STATE
The JSON state is the current user message text (`query`); the question instruction includes the candidate memory content. Batched IDs are m0, m1, and so on.
JEV · NOUL
YesNo
App workflow
- Screen memories
Batch candidate memories against current conversation context, then inject relevant results under the documented probability rule.
- Answer model question
The main model can invoke the judge tool for a constrained question.
- Fallback
When Jev is unavailable or confidence is low, retain the existing client behavior.
Why it is interesting
The same decision layer supports automatic memory selection and explicit on-demand questions, with silent fallback when Jev cannot answer.