Grok Bot 记忆不可靠:一周 31 条事实错 9 条
作者实测 Grok Bot 仅靠聊天记忆跑一周,31 条事实有 9 条日期或金额出错,主张把「第二大脑」放进 /workspace 文件而非聊天记忆。
somebody's NDA died so you could read this SpaceXAI sells Grok Bot as the coworker that remembers you, then wrote the catch into its own documentation and left the page up: a bot's memory keeps short summaries of your work, and the docs call it "not a substitute for the source of truth" Andrej Karpathy's second brain post had already pulled 22 million views, and the most saved build of it, 85,000 bookmarks, only works while Obsidian stays open on your laptop i ran one bot on chat memory alone for a week, counting what came back: 31 facts i had told it, 9 returned with the wrong date or amount, and every wrong one was a summary of a thread it never reopened a second brain on Grok Bot is worth building, and you can make it stop guessing before the first routine runs: - keep the brain in files in /workspace, not in chat memory: a file stays exactly as written, and every bot you add later reads the same folder - split facts in two: who people are and how you work live in files, dates, amounts and status get reopened at the source on every run - never edit raw/: everything lands there as it came and every page links back to it, so a wrong claim traces to its line in one click - keep routines to a few runs a day, since an hourly one can burn a week of your plan's bot work in a single day - give it a rules file it rereads before every task: an engineer at SpaceXAI has his bots review their rule book at 5 a.m. "so the same mistake doesn't happen twice" the cloud computer is also why this pays: the bot files your mail at 11 p.m. and hands you one page at 7:30 while the laptop is shut, exactly where the Obsidian builds stop my take, and it is the uncomfortable one: most second brains are a chat that forgets politely, and the fix was always one folder and one rule about where a fact lives bookmark this, the seven prompts that build the folder, the night filing, the morning page and the Sunday "what you decided and forgot" are written out in the course ↓
这条是引用着另一条写的:x.com/argona0x/status/2110045296902164532