It might just be me, but I'm confused: when I asked Claude to show me the memory of my project, it just spit out the directory of files within the project. Where/how do I specifically access the memory as distinct from the Project's files?
On a sort of related note, I find I'm using my project for 2 different types of task: 1. Generate a structured weekly meal prep report to spec, and 2. related "side conversations": stuff I liked and didn't like about particular recipes, now that I have this new to me ingredient let's brainstorm some other uses for it, snacks galore, general brainstorming on flavor profiles.... Is it possible to create a folder structure so that I can put different chat threads within the same project that do different things into different folders so I can keep track of which is which and find the ones I'm looking for more quickly?
1) have u turned on the memory on the settings? the project memory should show up above your master prompt section.
2) why don't you combine your two projects into 1? but this will require adjustment on the master prompt and project files too. Turn them into a health coach that help you achieve healthy lifestyle through meal prep that includes planning and brainstorming. Because later on if you want to reference different chats, Claude can read them as they are in the same project. And you can put comprehensive files inside your project too.
1) I only had the first memory setting on; I just now enabled "Generate memory from chat history."
How to get there wasn't intuitive to me, so in case this helps for anyone else: to find this setting first click edit project instructions, then at the top there's text includes a user preferences link. Click that link, go to capabilities (#6 in the drop down menu). From there in the third section there's 2 types of memory search and reference chats, and generate memory from chat history. Auto for me was the first one enabled, second one deactivated. To do all the cool stuff in this article seems like the second one needs to be enabled.
2) it is all in one project (for that very reason; the idea is for it to figure out what I like when both brainstorming/feedback and the actual plan documents live in the same project environment). I just wish I could file the threads within the project so that I had an easier time finding the ones with the specific plans vs. all the rest.
I love this. From the start, this is how I’ve worked with ChatGPT. But I had to brute force these systems into place, whereas Claude now seems built for them.
I built my assistant Kevin in ChatGPT so have stayed there for the most part, but I think I’m going to go play with Claude now and see if it’s time to move Kevin.
I do manage memory better than you describe in ChatGPT, though. I can manually delete anything from Kevin’s memory, and can add things with a simple command (add this to your permanent memory: “ “). I’ve written quite a bit about how I manage his memory and how it’s enabled me to train him. But I’m definitely going to experiment with Claude after this.
Nice! More things for me to experiment this week to improve how I use Claude. It is indeed difficult for us (humans) to always know what we want, especially when it's a creative task. "We always know more than we can say, and we will always say more than we can write down." (Dave Snowden from principles of managing knowledge). The only way is to iterate to clarity and you have presented a nice way to do this systematically with LLMs as an observation and sparring partner.
Yes i think we know things more instinctively but still find it hard to articulate things and AI can actually reveal that better as it can read the pattern. But to do it, it requires self reflection and deep awareness to evaluate their outputs and get the detail as much as possible. In the process, we can learn so much better with how we think in the first place!
Totally! This leads to us developing our inquiring skill, learning to be more curious, hopefully, building more open-mindedness to possible ways of seeing things.
Love this perspective on Claude's memory! It's super smart how you're leveraging AI for those viral hooks, 'cause LinkedIn really can be a minefield. I totally get what you mean about getting it past the 'fine but forgetable' stage. It realy highlights how the human prompt engineer is still key to finding that magic. So good.
As much as we try to create the 'perfect' output... the reality is we won't really know what 'perfect' looks like until we get terrible outputs. Basically, perfect prompts don't exist and iteration is part of the process but even so, AI can help with that part. Nice post, Wyndo!
love love love this process. Just got started with Claude skills a week ago so I'm bookmarking this. And as you said this is a huge addition to my AI voice workflow! Thank you for yet another great post ❤️
This is gold. The focus on friction as feedback is such a smart way to train AI while sharpening your own taste. It’s not just about making Claude better, it’s about making you more aware of what truly works.
Loved this guide, Wyndo. One thing I’m still figuring out is privacy and scope with Claude’s memory. How are you handling that in your own workflow, do you mostly lean on project‑level memory and keep things subject‑specific, or do you also let a more global memory build up and then use incognito for anything sensitive or edge‑case?
I use mostly project level memory. I don’t really open up on the global memory though. I like them to be neutral and focus on the work.
This is the problem with ChatGPT’s global memories. It can be double-edge sword where memories can help you but also trigger bias and enforce narrow perspective on some topics.
Super helpful, thank you, the trickiest part for me is drawing a hard line on what never belongs in persistent memory, especially around proprietary IP or anything client‑specific, and keeping that in local docs or short‑lived context instead.
Claude has tendency to skip some project files due to token efficiency or lack of context. Making it on the memory ensures it will never skip it and make it as priority with the same level as system/master prompt. So, I combine both of them for more effective output!
yes depending tasks. but I put everything I said on the example in this post into the memory.
when I encounter those cases, Claude can remember it and reference the right files as foundation to generate its outputs.
The limited characters is actually the input characters, but not the text inside the memory if you know what i'm saying here?
Try to input text to memory, you can't go details, but Claude can actually store huge amount of memory. I have project memory worth of long-form article now as it keeps adding more.
Ok. I think I get it now Wyndo. Thanks for sticking with me here to walk me through this. I think it’s a paradigm shift for how I use memory so I didn’t quite get it at first. But in tracking now. Thanks as always Wyndo
It might just be me, but I'm confused: when I asked Claude to show me the memory of my project, it just spit out the directory of files within the project. Where/how do I specifically access the memory as distinct from the Project's files?
On a sort of related note, I find I'm using my project for 2 different types of task: 1. Generate a structured weekly meal prep report to spec, and 2. related "side conversations": stuff I liked and didn't like about particular recipes, now that I have this new to me ingredient let's brainstorm some other uses for it, snacks galore, general brainstorming on flavor profiles.... Is it possible to create a folder structure so that I can put different chat threads within the same project that do different things into different folders so I can keep track of which is which and find the ones I'm looking for more quickly?
1) have u turned on the memory on the settings? the project memory should show up above your master prompt section.
2) why don't you combine your two projects into 1? but this will require adjustment on the master prompt and project files too. Turn them into a health coach that help you achieve healthy lifestyle through meal prep that includes planning and brainstorming. Because later on if you want to reference different chats, Claude can read them as they are in the same project. And you can put comprehensive files inside your project too.
1) I only had the first memory setting on; I just now enabled "Generate memory from chat history."
How to get there wasn't intuitive to me, so in case this helps for anyone else: to find this setting first click edit project instructions, then at the top there's text includes a user preferences link. Click that link, go to capabilities (#6 in the drop down menu). From there in the third section there's 2 types of memory search and reference chats, and generate memory from chat history. Auto for me was the first one enabled, second one deactivated. To do all the cool stuff in this article seems like the second one needs to be enabled.
2) it is all in one project (for that very reason; the idea is for it to figure out what I like when both brainstorming/feedback and the actual plan documents live in the same project environment). I just wish I could file the threads within the project so that I had an easier time finding the ones with the specific plans vs. all the rest.
1. Thanks for sharing this, i will update it on the post to include this process!
2. Hmm i think you can trigger it with: “remember the last chat about X?” And it will make reference from previous chat. Does this help?
Really actionable and helpful. Important to use memory for concise, always-on behavioral rules.
thanks Chintan, this is the way!
just catch the moments where the model misses your vibe, lock the fix, and the whole thing levels up!!
exactly, as simple as that!
I love this. From the start, this is how I’ve worked with ChatGPT. But I had to brute force these systems into place, whereas Claude now seems built for them.
I built my assistant Kevin in ChatGPT so have stayed there for the most part, but I think I’m going to go play with Claude now and see if it’s time to move Kevin.
I do manage memory better than you describe in ChatGPT, though. I can manually delete anything from Kevin’s memory, and can add things with a simple command (add this to your permanent memory: “ “). I’ve written quite a bit about how I manage his memory and how it’s enabled me to train him. But I’m definitely going to experiment with Claude after this.
curious to know if you like claude vs chatgpt better at managing and measuring impact of their memories. anyway, have fun!
Gold.
Nice! More things for me to experiment this week to improve how I use Claude. It is indeed difficult for us (humans) to always know what we want, especially when it's a creative task. "We always know more than we can say, and we will always say more than we can write down." (Dave Snowden from principles of managing knowledge). The only way is to iterate to clarity and you have presented a nice way to do this systematically with LLMs as an observation and sparring partner.
Yes i think we know things more instinctively but still find it hard to articulate things and AI can actually reveal that better as it can read the pattern. But to do it, it requires self reflection and deep awareness to evaluate their outputs and get the detail as much as possible. In the process, we can learn so much better with how we think in the first place!
Totally! This leads to us developing our inquiring skill, learning to be more curious, hopefully, building more open-mindedness to possible ways of seeing things.
Love this perspective on Claude's memory! It's super smart how you're leveraging AI for those viral hooks, 'cause LinkedIn really can be a minefield. I totally get what you mean about getting it past the 'fine but forgetable' stage. It realy highlights how the human prompt engineer is still key to finding that magic. So good.
Yes, human taste becomes more important than ever and we need to sharpen this more often.
As much as we try to create the 'perfect' output... the reality is we won't really know what 'perfect' looks like until we get terrible outputs. Basically, perfect prompts don't exist and iteration is part of the process but even so, AI can help with that part. Nice post, Wyndo!
Exactly!
There's no perfect prompt, there's only iterative process that keep evolving over time :)
Great post. It’s literally coding language 😀
love love love this process. Just got started with Claude skills a week ago so I'm bookmarking this. And as you said this is a huge addition to my AI voice workflow! Thank you for yet another great post ❤️
Have fun with Claude Skills! :)
This is gold. The focus on friction as feedback is such a smart way to train AI while sharpening your own taste. It’s not just about making Claude better, it’s about making you more aware of what truly works.
thats the secret of any great work!
Thanks for the good 😊
This is a really thoughtful one, Wyndo! I’m totally stealing some of your ideas 💡
Also, that promotional email cringe is so real, I have it too for some of mine lol.
My lesson is we just have to teach AI as much as possible and refine it over time. That's where it will get better.
Loved this guide, Wyndo. One thing I’m still figuring out is privacy and scope with Claude’s memory. How are you handling that in your own workflow, do you mostly lean on project‑level memory and keep things subject‑specific, or do you also let a more global memory build up and then use incognito for anything sensitive or edge‑case?
I use mostly project level memory. I don’t really open up on the global memory though. I like them to be neutral and focus on the work.
This is the problem with ChatGPT’s global memories. It can be double-edge sword where memories can help you but also trigger bias and enforce narrow perspective on some topics.
Super helpful, thank you, the trickiest part for me is drawing a hard line on what never belongs in persistent memory, especially around proprietary IP or anything client‑specific, and keeping that in local docs or short‑lived context instead.
yes, managing sensitive information is necessary here. Plus you can also remove unnecessary memories in case it doesn't serve you anymore.
Great article as always Wyndo.
Why not just put it on the project knowledge base and skip memory, if that is where is going in the end. Is it the compression you’re after?
I’m trying to understand why this step is necessary
Thanks Wyndo.
Claude has tendency to skip some project files due to token efficiency or lack of context. Making it on the memory ensures it will never skip it and make it as priority with the same level as system/master prompt. So, I combine both of them for more effective output!
So in this step:
Here’s what you can do: ask Claude to create the memory for you, generate it as a Markdown file, and upload it to your project knowledge base
Are you now exporting the memory to project knowledge ? So it will not be in memory anymore, freeing up space to put other stuff in memory?
If I underhand that correctly, then the memory you built is erased and goes to the project file.
Am I misunderstanding?
what goes into knowledge base is the detailed version of memory.
my example is my sales email. I put my sales email in memory but when it comes the detailed how to write sales email, it will refer to knowledge base.
so the sales email lives in two areas: memory (principle) and knowledge base (detailed framework).
because the problem with current memory input is limited, you cant go details unless you use Claude Code.
Oh! I thought you were clearing the memory.
So does that mean you have to manage memory depending on your task? Because as you said the characters are very limited.
This is the issue I’m having with Claude memory. That’s why I’m asking so much. Lol
yes depending tasks. but I put everything I said on the example in this post into the memory.
when I encounter those cases, Claude can remember it and reference the right files as foundation to generate its outputs.
The limited characters is actually the input characters, but not the text inside the memory if you know what i'm saying here?
Try to input text to memory, you can't go details, but Claude can actually store huge amount of memory. I have project memory worth of long-form article now as it keeps adding more.
Ok. I think I get it now Wyndo. Thanks for sticking with me here to walk me through this. I think it’s a paradigm shift for how I use memory so I didn’t quite get it at first. But in tracking now. Thanks as always Wyndo