Codex 用户为 GPT-6 Astra 建记忆文件夹，每次任务前重读

Mojunel.eth @mojunel原推发布于 2026.10.16

在 X 上打开原推

This Codex user built a memory folder GPT-6 Astra rereads on every job, so it's briefed before typing a word, says a 111,044-view post. 19 days in, the author claims the folder holds 165,000 leaves. I read the post and checked what of it is public. It's hand-made file memory that keeps a coding agent from starting cold, always on, while Codex's own memories ship switched off. → ENTRY: AGENTS.md tops a folder named memory, and Codex loads it at startup. → RULE: Read this folder before every task, then write what you learned and what broke. → SEED: A lease, a bank export and the manual of a device that never worked. → NIGHT: At 2am a routine works through every note tagged "later"; after incidents the model writes itself rules. → PRUNE: Every Sunday a 20-minute pass cuts dead branches, since the growing folder eats the weekly limit. What holds up > claimed: 41,000 files, 118,000 calls, 6,000 notes, per the author. > claimed: receipts from an hour to 6 minutes by the fourth run, per the same post. > checked: GPT-6 Astra shipped 3 September 2026, per secondary news sites. > checked: Codex loads AGENTS.md once at startup, up to 32 KiB, and only if it opens in that folder or its repo. > checked: the reread is a plain-text request, so it holds only while the model obeys. > checked: no page I read backs the $20-plan claim or the cheaper-helper reading. > not public: the folder, the logs, the rules file. No count can be checked. The method is buildable from public docs. The speedups are unverified. Copy it and your lease and bank export sit in plain files the model rewrites. Before trusting a 6-minute receipt run, time your own first and fourth pass. I laid the folder, both routines and every check into the terminal below ↓

这条是引用着另一条写的：x.com/mojunel/status/2106042827679408577