WHAT NEW IN X11.03 TUE

AI collections from X · 22 collections · 65,263 posts

jevVIDEO

fast-jev-compaction scores Claude Code context with Jev instead of summarizing

The plugin does not rewrite memory; it uses a small fast model to score each tool call and its output, cutting low-scoring ones and keeping the rest word for word, plus five things to check before installing.

Alex Stone
@a1exstoneArchived 2026.09.22Originally posted 2026.09.22
Alex Stone @a1exstoneOriginally posted 2026.09.22
Open original on X

CLAUDE CODE'S BIGGEST HIDDEN PROBLEM: IT FORGETS THE EXACT ERROR IT WAS FIXING. HERE'S THE PLUGIN THAT STOPS IT. when a session gets long, claude code squeezes the whole chat into a summary. file paths, error messages, the "never touch this folder" rule you gave it an hour ago. half of it doesn't survive. fast-jev-compaction: 5.8k stars, 319 forks. i read the whole readme. 5 things before you install it. it doesn't summarize anything. it looks at every tool call in your session and asks two questions: do we still need this call, and do we still need its output. a small fast model scores both. whatever scores low gets cut. everything else stays word for word, and your messages are never touched. -> CHECK your claude code version first. it needs 2.1.274+ and an early-access setting turned on. without it, nothing happens. -> WATCH which one you got. if the plugin fails or can't free up enough space, it quietly falls back to the old summary. the log tells you which one ran, so read it. -> RAISE the bar before you trust it. by default anything under 50% confidence gets cut. that's a coin flip deciding what your agent forgets. even the readme says a score is not a proof that something is safe to delete. -> COUNT the cost. the whole history is sent again with every batch of questions. a long session means several requests per cleanup, not one. -> IGNORE the demo video. it's a scripted mac app that never calls the api. the readme says it exists to be screen recorded. the tests use a fake model too. the only real check is running it on your own session. the idea is right: don't rewrite the memory, just throw out what's dead. i'd still give it a week on a side project before trusting it with real work. would you let a model you can't question decide what your agent forgets?

This post quotes another post: x.com/me_barnyx/status/2101976764779999381

x.com/a1exstone/status/2102083135777169795Archived 2026.09.22 10:13