用 Jev 给 Grok Bot 做记忆路由决策

Branko Trcek @Brankotrcek原推发布于 2026.10.24

在 X 上打开原推

THIS IS F*CKING GOLD your Grok Bot can have a second brain. the useful part is making yesterday’s work cheaper to use tomorrow the idea: keep the working chat short, store reusable knowledge in shared files, and turn proven workflows into skills. connect JEV to handle the small routing decisions between them your bot starts with what it already knows instead of charging you to figure it out again you set Grok Bot up like this: “before starting, check what we already know. retrieve what matters, investigate what changed, and save anything the next run can reuse” here’s the split: → shared files hold your preferences, decisions, and confirmed findings → JEV classifies the next step: reuse, refresh, or investigate → Grok Bot handles the research, writing, and uncertain cases → saved skills handle jobs you’ve already taught it give JEV the request and relevant saved evidence through an API tool. it returns a structured choice with confidence; your workflow uses that signal to route the task fresh research starts in 3 situations: → the relevant information is missing → a saved fact needs updating → new evidence contradicts the previous answer the guide’s worked example: before: 100,000 tokens per task after: 12,500 tokens per task same budget: 7 usable results → 70 memory supplies the background. JEV helps route the request. Grok Bot does the work that also changes how a team of bots operates: one bot records the finding, the next uses it. handing over a task shouldn’t restart the investigation the full setup: → ABOUT.md holds your goals, preferences, and constraints → BRIEF.md defines the current job and what counts as done → raw/ keeps original notes, documents, and source material → wiki/ stores reusable findings with sources and dates → RULES.md turns your corrections into standing instructions → skills/ holds workflows that already produced useful results → HANDOFF.md records completed work, open questions, and the next action don’t load the whole brain into every conversation. retrieve the useful section and leave the rest in storage paste this into Grok Bot ↓ “Rebuild my workflow around a shared second brain: → Review one recurring task and identify what I keep explaining or researching again. → Separate lasting knowledge from changing facts. Save reusable findings with sources and dates; check live information when the task needs it. → Keep the current brief and handoff short. Retrieve relevant passages instead of loading every file. → If JEV is connected, use it for defined routing decisions: reuse, refresh, or investigate. Escalate uncertain cases rather than forcing a choice. → Give each task one owner. Pass confirmed findings between bots so they don’t repeat the same research. → Preserve completed steps when something fails. Try one focused correction, then stop and report anything unresolved. Show me the proposed files and one test task before scheduling routines.” save this. the next run should inherit the work, not the entire conversation the article below breaks down 7 changes that keep this second brain useful while cutting repeated context, unnecessary checks, and failed retries ↓

这条是引用着另一条写的：x.com/greshamtrade/status/2112214741103825314