Your Claude still believes things about you that stopped being true months ago.
Old tone rules.
A handle you never use.
A format you dropped in September.
Every new chat, it acts on all of them as if nothing changed.
The problem is sleep.
Your Claude has been awake since the day you signed up, collecting notes and never cleaning any of them up.
On November 6, Anthropic published how they give Claude memory that works.
A big part of it is a process they call dreaming.
I tested the idea on my own Claude this week.
I have 38 memory files, and two of them were quietly wrecking my drafts.
If you’re new here, welcome. If you’ve been here a while, thank you for sticking with me :)
By the end of this issue, you’ll know exactly what your Claude remembers about you, and which of those memories are helping and which are lying.
You’ll also have a 10-minute routine that gets your Claude doing what Anthropic’s does: going back over its notes and waking up with fewer wrong ones.
By the end of this issue, you will get 20 copy-paste prompts for checking, fixing, and cleaning up Claude’s memory. They’re the same ones I use, and they’re free for everyone.
The week my Claude kept tagging a stranger
My X handle is @humzaakhalid, with a double “a.”
For weeks, Claude kept writing @HamzaKhalid in my CTAs.
That account belongs to someone else.
I’d catch it, fix it, and two days later it was back.
Somewhere in my memory, an old note said one thing and a newer note said another.
Claude had no way to tell which was true, so it guessed, and it guessed wrong.
Anthropic ran into the same thing in a video game.
Know someone who corrects Claude on the same thing every week? Send them this.
What Anthropic found
On November 6, Anthropic’s engineering team published a post called “The components of agent memory.”
You can read the whole thing here:
It’s long, so here are the three findings I’d act on. Every number below comes from that post.
1. One wrong note can trap Claude in a loop
Anthropic had Claude play Pokémon Red.
Claude wrote down the wrong location for the Pokémon Center near Mt. Moon.
Because of that one note, it kept misreading where it was, walked over a one-way ledge, and landed back on an earlier route, again and again.
Then Anthropic ran dreaming on its memory folder.
Claude read back through the transcript, compared the note to what had really happened, and fixed it.
On the corrected run, it went back into the cave and reached the next city.
A note is Claude’s summary of what happened. The transcript is the record. Dreaming checks the first against the second.
My handle problem was the same ledge.
2. Less memory went further
In the same game, about 10,000 steps in, Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge.
Fable 5 wrote nine files, plus one top-level file called global.md, and earned four.
Sonnet 3.7 recorded everything, but the game kept changing, so its notes kept going out of date. Fable 5 saved plans, strategies, and rules it had tested, and skipped what it could already see on screen.
This was one run per model, so treat it as a strong hint. Still, if your instinct is to make Claude remember more, their results point the other way.
This prompt fixed more of my drafts than any style guide. Pass it on.
3. Good memory says how it knows
After seven Poké Balls in a row failed to catch a Magikarp, Claude wrote itself a rule: if three or more balls fail in a row, change the timing. It also noted that the catch landed right after it did that.
I love that one. It’s a rule with its own proof attached.
The bigger test was Stainless, a real codebase of about 350,000 files. Claude built a wiki of it through dreaming. That wiki met 13% of the grading criteria after two dreaming sessions, and 48% after 128.
Testing Claude’s memory system took me two full days and more coffee than I’d like to admit. If this saves you even one of those days, buy me a coffee.
The Dream Routine
I run it in four stages, like a night of sleep.
Most people stop after the first.
Stage 1: Wake (see what Claude knows)
Open Settings.
Go to Your Claude.
Find Memory.
Check that memory is switched on.
Then find Your Memory and click on Manage.
Open the list and read every memory Claude has about you.
Read it slowly.
You’ll probably find at least one thing that isn’t you anymore. I found two.
Stage 2: Day (teach it with a reason)
When you tell Claude something worth keeping, tell it why too.
That’s the Magikarp lesson.
Instead of: “Remember my handle is @humzaakhalid.”
Say this:
Remember this: my X handle is @humzaakhalid, with a double "a."
Why: @HamzaKhalid and @humzakhalid belong to other people, and tagging them sends my readers to a stranger.
How to apply: check every CTA and tag for the double "a" before showing me the draft.The fact, why it’s true, and when to use it.
The “how to apply” line tells Claude what to do with the fact, so it doesn’t just sit there.
Stage 3: Dusk (review every good chat before it ends)
This is a small nightly dream. At the end of a chat that went well, paste this:
Before we close, read back through this whole chat like you're reviewing your day.
1. List what I taught you that should change how you work next time.
2. Include what went right, not only what I corrected.
3. Throw out anything that only mattered for this one task, and tell me why.
4. Check each one against what you already remember. If it clashes with an old memory, rewrite the old one instead of adding a new one.
Show me the list. Save nothing until I say yes.Only run this on good chats.
A bad chat teaches the wrong lessons.
Stage 4: Dream (the monthly clean-up)
Once a month, let Claude go over everything. Paste this:
Go through everything you remember about me, one memory at a time.
1. Find anything that's out of date.
2. Find any two memories that contradict each other.
3. Find anything you'd only ever need once.
4. Find any memory that states a rule without saying why.
Show me a table: the memory, KEEP, FIX or DELETE, and one line on why. For every FIX, write the new wording, including the reason. Don't change anything until I say yes.When I ran this, it flagged clashes I’d forgotten about.
One memory said never use emojis.
Another, from a month later, said one emoji is fine in a CTA.
Same with hashtags: one note said none, another said two at the end.
Claude had been obeying whichever one it read first.
Slow down for a second
Your Claude probably isn’t forgetting things.
It’s holding on to every old version of you and can’t tell which one is current.
In March you told it one thing.
In June you changed your mind.
You never deleted the March note, so Claude still has both, and as far as it can tell, both are you.
That’s how I ended up with the wrong handle and the lowercase rule.
Adding another “remember this” doesn’t fix it.
It just adds a third note to the argument.
Claude can only be as clear as the notes you leave it.
Most of the fix is deleting the old ones, and that’s the job dreaming does.
Two days, way too many coffee cups, and 38 memory files later, this routine is what survived. If it helped, keep the coffee coming.
Recap checklist
☐ Read every memory Claude has about you.
☐ Delete anything that isn’t you anymore.
☐ Teach new memories as fact, why, and how to apply.
☐ Save rules you’ve tested, not things Claude can already see.
☐ Run the Dusk review at the end of every good chat.
☐ Skip the review on bad chats.
☐ When a new rule clashes with an old one, rewrite the old one instead of adding another.
☐ Run the Dream clean-up once a month.
☐ Say yes to each change yourself. Never let Claude save without showing you first.
Start with Stage 1 tonight. It takes three minutes, and you’ll find something.
Know someone whose Claude is stuck on old memories? Send them this.
Your gift: 20 Memory Correction Prompts
These are the 20 prompts I use to keep my Claude’s memory clean.
They cover finding out-of-date memories, fixing clashing rules, rewriting vague notes with a reason, and cleaning up after a messy chat.
Copy one, paste it in, and say yes only to the changes you want.
We tested it in public. Now go use it in private.
I spent two days and a lot of coffee testing this so you could do it in ten minutes. If that feels like a fair trade, buy me the next cup.
P.S. Run the Dream clean-up tonight and reply with the strangest memory Claude had saved about you. Mine was a handle that belongs to a stranger.
I’ll share the best ones next week.















outstanding. About 100 records ended with 56 lines to manage but worth it. Mostly contradictions but a lot of lines flagged for deletion and keep right off the bat. Many thanks!