NewTreg Jev Memory
An opt-in Claude Code mod in Treg remembers lasting project preferences from answered prompts using Jev.
Added to Jevfast
NewHow it uses Jev
After a user-authored prompt receives an answer, the Claude Code mod splits the prompt into up to six sentences and sends them in one Jev request through Treg. It asks a Noul question per sentence about whether the sentence states a lasting project preference, rule, or fact. Sentences scoring at least 0.8 are saved verbatim after trimming to 300 characters and removing duplicates; saved lines are added to later session context unless the memory file is tracked by Git.
What Jev decides
Illustrative sentence preference judgmentExample answers · not a recorded Jev response · Source ↗
Question 1 · q1
YOUR APP
INSTRUCTION
Judge only this sentence for a lasting project preference, rule, or fact rather than a one-off instruction.
STATE
The state contains the prompt's extracted sentences, each tagged with an ID. Each sentence gets its own question in the same request. This is a source-derived request example, not an observed answer.
JEV · NOUL
YesNo
App workflow
- Collect an answered user prompt
The hook records prompts from supported user origins, then processes them after an answered turn.
- Split and judge sentences
The mod limits input to 4,000 characters and six sentences, creates sentence IDs, and asks one Noul per sentence in a single Treg call.
- Save qualifying preferences
The code applies the 0.8 threshold, trims saved text to 300 characters, removes duplicates, and writes accepted lines to .claude/jev-memory.md.
- Reuse and manage memory
At later session start the mod adds untracked memory lines to context. The /jev-memory command lists lines or removes one by number.
Why it is interesting
The mod adds a small preference memory loop around completed conversations. It keeps the model decision separate from deterministic thresholding, persistence, deduplication, and user controls to list or forget saved lines.