Claude Code memory system: Jev classifies what to remember
After each turn, Jev scores whether something is a lasting preference or a one-off task, saving only scores above 0.8 to a memory file, with 87 checks costing $0.0018.
CLAUDE CODE GOT A MEMORY SYSTEM COSTS $.0018 The trick is not stuffing more history into the context window. Claude Mod catches the end of every turn. Then a tiny classifier asks one question: > Is this a lasting preference or a one-off task? Example from the guide: In this repo, always prefix git branch names with sx/. → Jev score: 0.95 → save to .claude/jev-memory.md Build me a Tetris game from scratch. → Jev score: 0.03 → skip it The numbers are the part that makes this interesting: > 87 checks cost $0.0018 > $1 of free credits is reported as roughly 55,000 Jev calls > the check runs after the turn, so Claude does not wait for it > only a few saved lines enter the next session The architecture is tiny: Claude finishes → Mod catches the event → Jev classifies → code applies the 0.8 threshold → memory is saved or discarded. That is a much better memory primitive than dumping every conversation into the next context window.
This post quotes another post: x.com/thegreatest_sv/status/2080634616860320073