Small decisions. Interesting possibilities.Submit contentSubmit

How it uses Jev

NeuroLink sends the current request and eligible earlier user or assistant messages to its decision provider. It asks one Boolean relevance question per message. A confident “not needed” can remove a message before context compaction; uncertain answers and failed calls leave the existing compaction path in place.

What Jev decides

Illustrative context relevance requestExample answers · not a recorded Jev response · Source ↗
Question 1 · msg 0
YOUR APP
INSTRUCTION

Does this earlier message contain information still needed to answer the current request? A “no” means the request is about something else, or the message is small talk, an acknowledgement, or superseded.

STATE

The state contains the current request and indexed eligible earlier messages with role and text. The current request is capped at 4,000 characters, message previews at 1,200 characters; the six newest messages are protected, and at most 300 eligible messages are asked about in one pass.

JEV · NOUL
YesNo

App workflow

  1. Select eligible history

    The compaction helper protects the six newest messages and excludes tool-related, summary, skill, empty, and non-user/assistant messages.

  2. Build relevance questions

    It creates one Boolean question for each eligible message, with bounded message text and a namespaced key such as `msg__0`.

  3. Apply conservative answers

    Only confident “not needed” answers are drop candidates; one pass removes at most half of the eligible messages.

  4. Continue compaction

    If the call fails, yields no usable answer, or finds no confident drop, the helper returns no change so the existing positional compaction continues.

Why it is interesting

The relevance pass adds request-aware pruning ahead of positional compaction, with safeguards around recent turns, tool exchanges, and the amount removed. It turns a large conversation into a batch of small typed judgments.