Claude + Jev 做不会遗忘的第二大脑
Jev 负责保存/待办/跳过/复核的判定，模型只写 5–12 行笔记，号称 98.5% 决策准确率、0.3 秒。
EVERYONE'S QUIETLY RUNNING THIS. NOBODY'S POSTED ABOUT IT YET. Claude + Jev = a second brain that doesn't forget. Diogo Almeida, who built Jev, put it plainly: "The point of Jev isn't to make the model smarter in a conversational sense, but to make it completely deterministic at decision points. If you have a choice of N actions, the system should yield an exact choice without hallucinations or wasted tokens." here's what's actually happening under the hood, in order: 01 - jev decides what to keep. code enforces the rules. the language model is left with exactly one job: write the final note. 02 - every input gets a verdict the second it arrives - save, todo, skip, or flag for review - scored on privacy and on whether it'll still matter in 30 days. 03 - nothing crosses the filter unless it clears a confidence and value threshold. private details never leave your device. 04 - the model is only allowed 5 to 12 lines, one claim, no rewriting the source body. lossy compaction is banned by design. 05 - every new note gets checked against its neighbors - duplicates, relationships, contradictions - and linked automatically when it's actually relevant. 06 - 10 to 20 candidates get pulled via BM25 or embeddings. only 5, max, make it into the context window. 07 - a nightly pass proposes merges, flags what's gone stale, surfaces conflicts - and leaves the actual call to you. 08 - no rebuild required. it runs on your Obsidian vault, plugs into Claude Code, or works standalone as a memory framework. the number nobody's talking about: 98.5% decision accuracy, 0.30 seconds, memory built 6.6x faster than a frontier model doing the same job. here's the real twist. your agent was never forgetting because the model was weak. it was forgetting because nothing in the loop was ever deciding what deserved to be kept in the first place. full breakdown in the article below ↓
这条是引用着另一条写的：x.com/0xkerazcity/status/2102042835352428590