Last night's article explained why long sessions decay. Today you get the two cures: memory files that outlive every session, and a second agent — Hermes — whose defining feature is that it rewrites its own memory as it learns you.
Install Hermes on your machine (it runs on the same ChatGPT subscription), give your main project a memory file, and master the handoff ritual — so no project of yours ever depends on a single chat session again.
You need: your laptop · the attention-decay article read · your my-tool folder with spec.md · patience for one 10-minute install (the heaviest of the course — and the last).
From the landscape article you know the split: Codex builds with you at the desk; Hermes learns you over time. Hermes keeps plain-text notes about your projects and preferences — and rewrites them itself as you work. Correct it once, and the correction is written down; next week it starts smarter. Two agents also unlock the professional pattern you'll use in week 4: one builds, the other reviews.
Windows only: install the Linux layer (WSL)
Hermes runs in a small Linux environment inside Windows — Microsoft's official feature, one command to add. Open PowerShell as administrator (Start → type “powershell” → right-click → Run as administrator):
wsl --installInstalling: Ubuntu … this takes a few minutesRestart your computer to finish.
Restart when asked. After reboot, a window finishes the Ubuntu setup — it asks you to invent a username and password (short, memorable; the password doesn't show while typing — that's normal).
No extra layer needed — open Terminal and continue with step 2.
Step done when: typing wsl in PowerShell drops you into a green-tinted Linux prompt (type exit to leave).
Install Hermes and sign in with ChatGPT
Inside WSL (Windows) or Terminal (Mac), run the official installer, then start Hermes and pick the ChatGPT sign-in:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash… installs Python and Hermes, no admin needed …hermesChoose a provider → Other providers → OpenAI OAuth (ChatGPT)A browser opens — sign in with YOUR ChatGPT account and approve.
Same subscription, second agent — Hermes uses your plan's included quota, no separate bill.
PS C:\…).Step done when: Hermes greets you in the terminal and answers “what can you do?”.
Watch it learn — the two-session experiment
Tell Hermes two personal preferences and one correction, then restart it and check what survived:
Three things to remember about me: I prefer answers in short bullet points; my main project is [one line from your spec.md]; and I work in [your profession]. Save these to your memory, then show me what you wrote.Read what it wrote — real files, in plain language. Now type exit, start hermes again, and ask: “what do you remember about me and my project?”
✦ A new Codex thread starts with an empty desk — old conversations don't follow you into it by default. Hermes just carried your facts across that gap, because its memory is files it maintains itself.
Fair note: Codex has an optional Memories setting (off by default — turn it on, it helps). It saves useful context from past threads in the background, but Codex decides what to keep. Hermes puts memory at the center: plain files you can open, read, and correct. And for project decisions, the next block's file is the memory you control.
Step done when: a fresh Hermes session recites your three facts.
Last night's article gave you the theory: chat history decays, files are re-read fresh every session. The practice is one file: AGENTS.md in your project folder. Codex reads it automatically at the start of every thread; Hermes reads it too. Decisions written there are permanent, front-of-mind knowledge — for any agent, forever.
How do lines get into memory? Two ways: you dictate them (“save this as a standing rule”), or — with Hermes — the agent writes them itself when it learns something worth keeping, like a correction you made. Either way, only durable things belong there:
Click each card: does this belong in the memory file, or is it just chat?
Give your main project its memory
So far we talked about Hermes's memory of you — your preferences, across all your work. A project needs a different memory: what this tool is, what's already decided, what the rules are. That memory lives in the project folder itself, in AGENTS.md, so ANY agent that opens the folder — Codex today, Hermes tomorrow — starts with the same facts. Your capstone gets one now:
Create AGENTS.md for this project:
1. The goal in one line (from spec.md).
2. A "Decisions" section — what's already decided.
3. Three standing rules: no new libraries without asking me; never touch files outside this folder; show me a plan before any change bigger than one file.
Under 25 lines. Show me the file when done.Then verify every line — this file is your project's constitution; a wrong line here repeats its mistake in every future session.
Step done when: AGENTS.md exists and you've personally read and approved each line.
The handoff ritual — kill a session without losing anything
Two prompts you'll use daily from now on. Ending a work session:
We're stopping. Update AGENTS.md with a "## Next session" section: what's DONE, what we DECIDED today, what's NEXT, what to AVOID. Under 12 lines.Starting the next one:
Read AGENTS.md and spec.md, then tell me the current plan in 5 bullets. Don't build anything yet.Rehearse the full cycle now: handoff → close the thread completely → new thread → read-back. Compare the 5 bullets with reality.
✦ The session died; the project didn't notice. From today, a foggy session costs you two minutes, not an afternoon.
Step done when: a brand-new session recited your project correctly with no explanations from you.
AGENTS.md holds rules and next steps — enough for a small project. When a project grows past a few files, professionals add a second file: summary.md, the full project guide. It answers, for an agent with zero context: what is this, what state is it in, how do I run and check it, where does everything live, and what breaks if touched carelessly. Updated at the end of every serious session — so any agent, any day, can pick the project up cold.
Create or update summary.md — the guide for an agent that has never seen this project. Cover: what this is (3 lines), current state (what works, what's unfinished), how to run and verify it, a map of the files (one line each), and landmines (what breaks if changed carelessly). Verify every claim against the real files — never write from memory. End with "Last updated [date] — [what changed]".Make the memory system yours: one real handoff, one observed Hermes learning moment.
homework complete — mark the day done below
Tick when the homework above is finished — it syncs to your account.
Attention decay — the decision drowned under newer tokens. Treatment: handoff prompt, confirm the decision landed in AGENTS.md, kill the session, read-back in a fresh one. Two minutes, complete recovery.
Codex reads memory files you maintain (AGENTS.md); Hermes also maintains its own. Both persist across sessions — but Hermes folds your corrections into its notes automatically, so it compounds without you doing the librarian work.
In: the goal, standing decisions, safety rules, what's next. Out: one-off requests, moods, today-only context. If the file stops fitting on one screen, it's becoming a junk drawer — decisions and rules only.