How to Build a Memory That Never Forgets with Opus 5.5 (Full Course)
文章目录
Think about the last great book you read.
Now tell me the three most important ideas in it.
You can't.
Neither can most people.
The research on this is old and brutal. Within days, most of what you learn is gone. Not some of it. MOST of it.
Every podcast on your commute. Every article you nodded along to. Every conversation where someone told you something important.
In one ear. Out the other.
You're not bad at remembering. Your brain was never built to store everything. It was built to think.
So stop asking it to do both.
In this course I'll show you how to build an external memory with Claude Opus 5.5. A system that captures everything you read, watch, hear and learn, turns it into memories you can actually use, and hands them back the exact moment you need them.
Not a notes app. Not another subscription. A folder of plain text files that you own forever, with Opus 5.5 running it for you.
By the end you'll be able to ask:
"What did that book say about pricing?" "What did Sarah tell me about her startup in March?" "What do I actually know about sleep?"
And get a real answer, with the source, in seconds.
Let's build it.
Why every notes app failed you
You've probably tried this before. Notion. Evernote. Apple Notes. A notebook. Highlights in your Kindle that you never looked at again.
They all fail for the same three reasons.
1. Capture is too slow.
If saving something takes more than ten seconds, you won't do it. So the best ideas never make it in.
2. Nothing ever comes back out.
You save 1,000 things and open 10. A notes app is a place ideas go to die. There's no system pulling the right note back to you at the right time.
3. You save quotes, not understanding.
A highlighted sentence means nothing six months later. Without the context, without why it mattered, without how it connects to what you already know, it's just noise.
Here's the shift that fixes all three.
You capture. Opus does everything else.
You throw raw material in. Opus reads it, pulls out what matters, writes it as a memory in your words, links it to what you already know, and brings it back when it's relevant.
Your only job is to feed it.
How the system works
The whole thing runs on five moves:
- CAPTURE → anything goes into inbox/ in seconds
- PROCESS → Opus turns raw material into memory cards
- CONNECT → every card links to related memories
- RECALL → ask any question, get answers with sources
- RESURFACE → old memories come back before you forget them
And it lives in one folder on your computer:
memory/ CLAUDE.md the rules Opus follows every time inbox/ raw stuff goes here, unprocessed sources/ the original material, saved untouched memories/ one card per idea, in plain language people/ one file per person you want to remember daily/ a short log of what you learned each day .claude/commands/ your 4 memory commands
Everything is markdown. Plain text files. They open in any editor, sync with any cloud drive, and will still open in 30 years.
If you use Obsidian, point it at this folder and you get a visual graph for free. But you don't need it. Claude Code is the engine.
Step 1: Create the folder
Install Claude Code. Then open your terminal:
mkdir memory cd memory mkdir inbox sources memories people daily .claude .claude/commands touch CLAUDE.md claude
Inside Claude Code, type /model and select Opus 5.5.
That's your memory engine, live.
Step 2: Write the rules of your memory
This file is the most important part of the whole system. Claude Code reads CLAUDE.md at the start of every session, so these rules shape every single memory it writes.
Paste this into CLAUDE.md:
My Memory System
You manage my external memory. Your job is to make sure I never lose anything worth knowing, and that I can find it again in seconds.
Folders
- inbox/: raw captures. Process these, never leave them. * sources/: original material. Save it UNTOUCHED before you do anything else. * memories/: one card per idea. * people/: one file per person. * daily/: one log per day, under 20 lines.
Memory card rules
- ONE idea per card. If a source has 5 ideas, make 5 cards. * Write the idea in plain language, like explaining it to a smart friend. Not a quote dump. * Every card links to its source in sources/. * Every card links to at least 2 related cards using [[card-name]] if they exist. * File names: short and lowercase, e.g. "pricing anchors.md"
Recall rules
- Answer ONLY from my memory files. * Cite the card and source for every claim. * If my memory doesn't contain the answer, say: "Not in your memory yet." Then list what's closest. * Never mix in outside knowledge without marking it [OUTSIDE].
Safety
- Never delete or rewrite anything in sources/. * Never delete a memory card. Mark outdated ones with status: outdated and link the newer card.
Three rules here do most of the work.
"Save it UNTOUCHED before you do anything else." Your original material is always safe. If Opus summarizes something badly, the source is still there.
"ONE idea per card." This is what makes recall work. You can't link, search or resurface a page that contains twelve different ideas.
"Not in your memory yet." Without this, Opus answers from the internet and you can't tell which parts came from your own life. With it, your memory stays honest.
Step 3: Design the memory card
Every idea becomes a card that looks like this. Save it as memories/_template.md so Opus copies it:
type: memory source: [[sources/book the mom test]] captured: 2026-11-08 last_reviewed: 2026-11-08 review_count: 0 tags: [customer research, startups]
People lie about the future, not the past
When you ask customers "would you use this?", they say yes to be nice. Ask about what they actually did last time they had the problem instead. Past behavior is real. Future promises are polite fiction.
Why it matters to me
I've been validating ideas by asking "would you pay for this?" That's the exact question this warns against.
Connected to
- [[talk is cheap commitments are real]] * [[customer interview questions]]
Original wording
"Opinions are worthless." (ch. 1)
Look at the section called "Why it matters to me."
That's the difference between a quote and a memory. Six months from now, the idea alone is easy to forget. The idea plus why it mattered to YOU sticks.
The last_reviewed and review_count fields power resurfacing later. Don't skip them.
Step 4: Build your 4 memory commands
Commands turn your whole system into four short words. Each one is a markdown file in .claude/commands/. Whatever you type after the command replaces $ARGUMENTS.
Command 1: /remember
Create .claude/commands/remember.md:
Process everything in inbox/ into my memory.
For each item:
- Save the original, untouched, to sources/ with a clear file name and the date.
- Pull out every idea worth keeping. Skip filler.
- Write ONE memory card per idea in memories/ using memories/_template.md.
- Fill in "Why it matters to me" using what you know about my current projects from my recent daily notes. If you can't tell, write "[ADD YOUR WHY]".
- Link each card to at least 2 existing related cards.
- If a person is mentioned, update their file in people/.
- Add a one line entry to today's file in daily/.
- Empty the inbox.
Then show me: how many cards you created, their titles, and anything you weren't sure about.
Command 2: /recall
Create .claude/commands/recall.md:
Answer this from my memory only:
$ARGUMENTS
- Search memories/, people/ and daily/ for everything relevant.
- Give me a short, direct answer.
- Cite every card and source you used.
- If cards disagree with each other, show me both and say which is newer.
- If the answer isn't in my memory, say "Not in your memory yet" and list the closest things I do know.
Command 3: /resurface
Create .claude/commands/resurface.md:
Pick 5 memory cards for me to review today.
Choose cards I'm about to forget, using last_reviewed and review_count: * review_count 0: due 1 day after captured * review_count 1: due after 3 days * review_count 2: due after 7 days * review_count 3: due after 21 days * review_count 4+: due after 60 days
For each card:
- Ask me one question that tests the idea, without showing the answer.
- After I answer, show the card.
- Update last_reviewed to today and add 1 to review_count. If I got it wrong, reset review_count to 1.
This is spaced repetition, the same principle behind flashcard apps used by medical students, except Opus writes the questions from YOUR memories.
Command 4: /review
Create .claude/commands/review.md:
Run my weekly memory review.
- List every card created this week, grouped by theme.
- Find 3 connections between cards I haven't linked yet and add the links.
- Find any cards that contradict each other and show me.
- Find any [ADD YOUR WHY] cards and ask me to fill them.
- Tell me the 3 biggest ideas I learned this week in one sentence each.
- Write the summary to daily/ as this week's review.
That's the whole interface.
/remember feed it /recall ask it /resurface keep it /review connect it
Step 5: Make capture take 10 seconds
The system only works if capture is effortless. Here's how to feed it from everywhere.
Articles and web pages
Copy the URL, open a file in inbox/, paste it. Or just paste the text. Opus reads either.
You can also drop it straight into Claude Code:
save this to inbox/: https://example.com/article
Books
Kindle keeps your highlights in a file called My Clippings.txt on the device. Plug in your Kindle, copy that file into inbox/, and run /remember. Every highlight becomes a card linked back to the book.
For physical books, snap photos of the pages you marked and drop them into inbox/. Opus can read the text from images.
Podcasts and YouTube
Grab the transcript (most YouTube videos have one under the description, and many podcast apps show them now) and paste it into a file in inbox/. Add one line at the top saying who it's from and why you saved it.
Conversations
Right after a call or a coffee, dump it in raw:
inbox/coffee with sarah.md
Sarah, runs a fintech startup, 12 people. Raised seed in Sept. Struggling to hire engineers. Said their best hires came from open source contributors. Wants intros to designers. Has a dog named Mochi.
Opus turns that into a people/sarah.md file plus a memory card about hiring from open source. Next time you see her, /recall what do I know about Sarah and you'll remember Mochi.
Your own thoughts
Ideas in the shower count too. Text yourself, then paste into inbox/ at the end of the day. Your own thinking is the most valuable memory of all.
Step 6: Your daily rhythm
Here's what using this looks like day to day. Total time: about 10 minutes.
Through the day: drop anything worth keeping into inbox/. Don't organize. Don't think. Just drop.
Evening, 5 minutes: run /remember. Skim the cards it created. Fix any "Why it matters" that missed the point.
Morning, 5 minutes: run /resurface. Answer 5 questions with your coffee. This is the step that moves knowledge from the folder into your head.
Sunday, 15 minutes: run /review. See what you learned. Find new connections.
That's it. Ten minutes a day and you'll remember more than people who spend hours rereading.
Step 7: Ask your memory hard questions
This is where it gets fun. After a month of capture, your memory starts answering questions you never thought to ask.
Try these:
/recall what do I actually believe about pricing?
/recall what have I learned about sleep, and where does the advice disagree?
/recall who do I know that could help with hiring?
/recall what ideas from this month connect to the project I'm working on right now?
/recall what did I think about remote work a year ago, and has that changed?
That last one is wild the first time you run it. Your memory shows you how your own thinking evolved, with dates and sources.
And because of the "Not in your memory yet" rule, you also see your gaps. That's a reading list built from your own blind spots.
Step 8: Make it remember you better over time
Add a file called me.md in the root of your folder:
About me
Current projects
- [Your main project and what stage it's at]
What I'm trying to learn
- [Topic 1] * [Topic 2]
How I like answers
- Short. Bullet points. Examples over theory.
Then add one line to CLAUDE.md:
Before writing "Why it matters to me", read me.md.
Now every memory card gets written through the lens of what you're actually working on. A chapter on negotiation gets tied to the deal you're closing. An article on habits gets tied to the routine you're building.
Update me.md once a month. It takes two minutes and makes every card sharper.
Step 9: Keep it private and safe
Your memory will hold things you care about: conversations, plans, personal notes. Treat it that way.
Keep it local. The folder lives on your machine. Sync it with a cloud drive you trust if you want it on multiple devices.
Back it up. It's just text files. Copy the folder somewhere safe weekly, or put it in a private git repository.
Keep secrets out. No passwords, no ID numbers, no financial account details in your memory. It's for knowledge, not credentials.
Be careful with other people's information. Notes about friends should be the kind of thing you'd be fine with them seeing. "Has a dog named Mochi" is great. Gossip is not.
The 5 mistakes that break the system
1. Organizing instead of capturing.Don't build a perfect folder structure. Drop things in the inbox and let Opus file them.
2. Saving everything.Not every article deserves a card. If you wouldn't tell a friend about it, skip it. Quality beats volume.
3. Skipping the "why."A card without "Why it matters to me" is a quote. Quotes get forgotten. Fill it in.
4. Never resurfacing.Capture without review is just a nicer graveyard. The 5 minute morning /resurface is what makes knowledge stick.
5. Letting it answer from the internet.Keep the "Not in your memory yet" rule. The whole value is knowing what YOU learned, from WHERE.
The copy and paste starter kit
Everything in one block to set up in five minutes:
mkdir memory && cd memory mkdir inbox sources memories people daily .claude .claude/commands touch CLAUDE.md me.md memories/_template.md touch .claude/commands/remember.md touch .claude/commands/recall.md touch .claude/commands/resurface.md touch .claude/commands/review.md claude
Then paste in:
-
The rules from Step 2 into CLAUDE.md
-
The card from Step 3 into memories/_template.md
-
The four commands from Step 4 into their files
-
Your projects and interests into me.md
Drop your first five things into inbox/ and run:
/remember
Watch your first memories appear.
What the first 30 days look like
Here's what to expect so you don't quit in week one.
Week 1: It feels like extra work.
You're building the habit of dropping things into the inbox. The memory is small, so /recall mostly says "Not in your memory yet." That's normal. Keep feeding it.
Week 2: The first surprise.
You'll run /resurface and get a question about something you saved ten days ago. You'll remember it. That's the moment the system clicks, because a month ago that idea would have been gone.
Week 3: Connections start appearing.
Your /review links an idea from a podcast to a chapter from a book and a note from a conversation. Three sources you'd never have put together on your own, sitting side by side.
Week 4: It starts answering real questions.
You're writing a post, prepping a meeting or making a decision, and you ask your memory first. It hands you what you learned, who told you, and where it came from.
By the end of the month you'll have somewhere between 100 and 300 cards, depending on how much you feed it. That's already more than most people retain in a year.
And it only gets better from here, because every new memory has more old memories to connect to.
The real shift
Most people consume more information in a week than their grandparents did in a year.
And they keep almost none of it.
The problem was never input. You read plenty. You listen to plenty. You talk to plenty of smart people.
The problem is that nothing stays.
This system changes that. Every book you finish, every podcast you hear, every conversation that matters becomes something you keep and build on.
A year from now you'll have thousands of memories, all connected, all searchable, all yours.
And when someone asks what you learned from that book, you won't say "I don't remember."
You'll know.
If you want to master AI and build systems like this, follow @cyrilXBT. I share every setup and prompt I use, free.
If this helped, like and repost it so someone on your timeline stops forgetting everything they read.
沿着这个主题继续读
从一篇文章,找到相关的教程、经验和阅读路径。



