Where Anthropic Kind Of Missed The Mark On Memory

What Was I Working On Last Tuesday?

To show how much this persistent memory system does for my Claude Code, I typed in one question: what was I working on last Tuesday? I asked for the answer as a simple, interactive document on my desktop, in list format, with a section for my top ten open tasks with checkboxes and every client name anonymized.

That Tuesday was August 4th, so the document had to detail exact memory from that date.

The document it came back with showed the tasks and jobs I was actually working on. It even logged its own misses. It had once proposed a Windows toast notification and told me to look for it bottom right, and those fail silently on this machine. That was already recorded back in July.

The interactive checklist landed right on my desktop, and when I opened it, it was exactly what I'd asked for. It remembers things by context and by date, plus anything I tell it to remember, and it indexes each one into a category so it comes back up when I'm working on similar jobs.

Each Project On Its Own

Anthropic has rolled out some memory recall in Claude's web and desktop clients. It writes its own entries from your chats as you go, and we want that. We also want it to search your past conversations when you ask it to, so you don't spend a significant amount of time reprompting for a job you're working on. It currently holds your role, your preferences and your ongoing work too.

It also gives each project its own separate memory, and this is where Anthropic kind of missed the mark. In the desktop and web versions, memory is held within each project on its own. If you're working inside one project, it won't have context from your other projects.

The second brain records what happened and what was said, and it records the entire process you went through, so it learns each time you start a new project. Every project, decision and lesson is tied to the others.

These records are kept as plain markdown files on your own computer, so you can back them up if you want and alter them yourself. On Claude's servers, you're relying on Anthropic to store your information.

The Learnings Folder

There are three things this memory system does that I think are a superpower for workflows and for setting up operations, for yourself or for your clients. The first, and the one I always recommend people set up, is a learnings folder.

These are lessons. Anytime I'm working on a project and Claude makes a mistake, gives me a poor outcome, or does something I didn't want or didn't anticipate, I go into my terminal, tell it to correct the mistake and record the lesson to our learnings folder. The lesson gets linked into its operating system so it doesn't make the same mistake again.

How many times have you prompted something in Claude and spent a lot of time reprompting the same thing to get a similar outcome? The learnings folder takes away that need to reprompt and correct, and that's what lets your work compound.

A Fraction Of The Time

Compounding is the second thing. Say you're building a dashboard, or doing something complex like an automation with ten different steps. The first time through, it takes you and Claude a lot more time. Getting the steps in order and connecting the different data sources takes a significant amount of time, and then you still have to test run the automation you've built.

With the persistent memory system set up, it remembers the entire process from the very start to the very finish, and you'll probably add some lessons along the way. The next time you build something similar, it takes a fraction of the time it took the first time, because Claude remembers the process: how you connected those data sources and how to test and do dry runs to make sure the automation is working properly.

I'm faster on my side too. I've already been through the process once, and I've got a digital brain that remembers all of the steps for me, so I don't have to.

The Daily Log

Then there's the daily log, which I think is essential to have. It opens at the start of each session and records detailed summaries, and at times exact transcripts, of what we're talking about, so there's a running log of what happened on each day.

That log is what let me create the document I asked for at the start. Claude remembers what you do every single day, and every workflow you've ever worked on is recorded in this memory.

A Little Afraid Of The Power It Has

Because Claude Code runs this memory system directly on your machine, you have 100% autonomy over what it can create, delete, modify and adjust on your PC. You can tell it to drop specific files in specific locations, which keeps your files neat and tidy.

When I first started doing this, I was blown away, and a little afraid of the power it has. I've fed it so much context about myself: my strengths and my weaknesses, my goals, my ambitions, my desires. It knows and remembers a lot about me because I wanted it to know those things.

I use it as an assistant. It loads its own context every session, so it knows what I'm working on and what I want to be moving towards.

It prevents you from starting at complete zero every single time.

The Prompt I Keep Improving

The second brain blows the native chat-based memory in the desktop and web clients completely out of the water.

This persistent memory system has gone through probably five or six iterations, and I keep improving it. I keep it as a live deliverable: a master prompt that sets it up for you, which I've tested and retested and made as simple and easy as I can.

This isn't going to work with Claude Desktop or Claude Cowork. It works inside a terminal, with Claude Code downloaded on your machine, so the brain can connect to all of your workflows. If you don't want to use Obsidian, a program you can connect to your Claude Code, the prompt works just as well without it. I'd still encourage you to set it up in Obsidian, because it's really fun, and it will show you the holes in your own memory system.