I Built a Memory System for Claude. Here’s How It Works.
What happens when Claude arrives at a client session already knowing the brief, the SoW, and what success looks like.
Every week, writing social posts for a client, I’d stare at a blank document.
The posts needed to fit the themes we’d agreed. What were the themes again? They were in the SoW. Somewhere. I could dig out the PDF, scroll through it, find the relevant section - or I could just ask Claude.
Except Claude didn’t know either. Every session started from scratch.
That friction, week after week, is what made me build the memory system.
If you haven’t set up Claude Cowork yet, the setup guide covers the foundations. This article is the next layer - specifically, how to build a client memory that means Claude arrives at every session already briefed.
The problem
Claude has no memory between sessions. Whatever you told it last week - your clients, your agreed scope, your goals - is gone. Every session is a blank sheet. It’s different to ChatGPT in that way.
For most tasks, that’s manageable. You brief Claude, it responds, the session produces something useful.
For client work, it’s a different problem. Client work isn’t one conversation. It’s dozens - calls, briefs, revisions, check-ins. It carries context: what was agreed in the SoW, what success looks like for this client, how they want their brand to sound. You hold all of that. Claude holds none of it.
So every session, two choices: spend time briefing from scratch, or risk doing work that missed the brief.
What changed the stakes
A client needed article titles approved before she went on leave the next morning.
If they missed the brief and needed revision, I wouldn’t hear back for a week. A week of blocked content means a month of work that I wouldn’t be able to bill at the amount we’d agreed.
I had one shot and my forecast monthly income told me that I had to get it right.
Claude didn’t need briefing. It already knew the SoW terms, the success objectives, the brand voice, what had come up in every call, the keywords they were targeting.
So I gave Claude the task and the titles came back. They were approved the same morning.
The weekly blank-page problem and the one-shot titles session are the same problem at different stakes. Both required Claude to know the client. Without the memory system, it couldn’t.
The client layer
In the setup guide, step three covers building an About Me folder: a set of files Claude reads at the start of every session - who you are, how you work, writing rules, and a memory log of decisions and project context.
The client layer is an extension of that folder.
Instead of one memory file tracking general context, there is a file for each active client. Claude reads it at the start of any session involving that client and arrives already briefed.
No code and no technical setup. Claude creates and updates the files itself based on what comes up in conversation.
What the client files hold
For each active client, the memory holds four things.
Transcriptions of our calls. Every conversation Granola records gets processed and filed. Claude knows not just what was agreed, but what was discussed, what the client pushed back on, and what came up that didn’t make it into the formal brief.
The SoW and contract terms. Scope, deliverables, billing structure, what’s in and what’s out. Claude can reference the exact agreed terms without me digging out the original document.
Success objectives. What does good look like for this client? What are they actually trying to achieve? This is the layer that separates work that fits from work that is generic.
Brand language. How they talk, what they avoid, the specific terms that matter to them. When Claude knows this from the actual contract and call transcripts rather than a brief summarised from memory, the quality difference shows.
What it changed
The social posts are the clearest example of the day-to-day difference.
I open a session. I say I’m writing this week’s posts for the client. Claude already knows the agreed themes, the brand voice, the topics from the last call. I don’t say any of that. I just write.
The blank page problem hasn’t gone away - writing is still writing. But the uncertainty about what was agreed has. That particular friction, every week, is gone.
The titles session shows what that means at higher stakes. Full client context meant first-time approval. First-time approval meant the month’s content could proceed. The month’s content proceeding meant I could bill what we’d agreed.
A memory system sounds like a time-saving tool. At client-service level, the difference shows up in what you can bill.
How to set it up
If you’ve followed the setup guide, you already have an About Me folder and a global instruction telling Claude to read it at the start of every session. The client layer builds on that in three steps.
First, create a clients folder inside your Cowork workspace and add a file for each active client. Tell Claude what to put in it: the client name, the agreed scope, the objectives, and any brand notes you have to hand.
Second, update your global instructions to tell Claude to read the relevant client file at the start of any session involving that client. One line is enough.
Third, after every client call, tell Claude to update the file with what was discussed. If you use a transcription tool like Granola, ask Claude to process the transcript and pull out what’s relevant.
The files improve every time you use them. After a few sessions, Claude knows the client almost as well as you do.
If you haven’t set up the foundation yet, the setup guide is the place to start. The client layer works best once the basics are solid.
The blank page used to mean two things: the writing ahead, and the uncertainty about whether I had the brief right.
Now it just means the writing.
If this is useful, there’s more like it every Saturday. Forward it to someone who’s still digging through old PDFs to remember what they agreed with a client.
Four Hour Freedom is a newsletter about building sustainable work around your life. If this was useful , I’d love it if you subscribed to the newsletter.
Failing that, you can Buy me a coffee (I really appreciate it!) or follow me socially on: Bluesky | Substack | Instagram