用 Jev 做项目记忆筛选，降低 Opus 5.5 重复开销

Branko Trcek @Brankotrcek原推发布于 2026.10.15

在 X 上打开原推

CUT YOUR OPUS 5.5 BILL WITH JEV AND A PROJECT BRAIN save the decisions, fixes, and context your agent already worked through, then let Jev select what the next task needs so you spend less on repeated explanations and unnecessary investigations 3 layers: files store the memory, Jev selects what matters, and Opus uses that context to continue the work write what matters: have Opus record the decisions, constraints, and working fixes, then use Jev to classify which notes belong in long-term project memory keep it clean: let code catch duplicates and use Jev to flag notes that may conflict with newer decisions so Opus can review and update them recall what’s relevant: give Jev the current task and candidate notes, score their relevance, and pass the selected context to Opus before another investigation begins if you fixed an authentication bug last week, today’s login task should surface that fix, the reason behind it, and the files it affected while unrelated design notes stay in storage you’re building a record of solved problems that the agent can consult before spending time and tokens working through them again the loop: work → record → consolidate → Jev selects context → Opus continues → update memory start with one recurring task, keep the useful notes, and compare the cost of completing it with and without that memory then bring Jev into task routing and retries so Opus gets a clear reason for every new call, with the complete setup laid out in my Opus 5.5 + Jev article below ↓

这条是引用着另一条写的：x.com/0xnineinch/status/2108604801411879139