Six months ago, I migrated all of my marketing work to Claude Code and didn’t think too much about its memory. You see, it starts as a tabula rasa, and as you do your work, you ask it to remember things. Then, as things progress, it sometimes decides to save something on its own. And over the days and weeks, the marketing OS gets a little smarter. Is it becoming sentient? Not a chance, but it does give you a hidden superpower.
It turns out that the Claude memory file is only a small part of what I’d consider memory. When I look at what Claude uses to create a data sheet, prep a podcast interview, or build a competitive brief, the curated memory facts are a tiny portion of what it uses. I have about 197 of them saved across my projects. Underneath those memory snippets sit 2,323 files in my consulting repository, another 576 in the TinyTechGuides repository, and 43 skills that specify workflows on how I want specific jobs done. Those saved facts work like an index, and the files themselves are the brain.
So six months in, I started asking a different question. If the memory feature is the least important kind of memory I have, what is doing the real work? Everything else in the project folder is. Each meeting note, customer transcript, and case study is already stored in the system, so I never load them by hand again. The same goes for the skills, the project rules, and the MCP connections to my live data. The lesson from half a year of running marketing this way is that the memory worth having is the entire body of work you accumulate, and it builds if you have a systematic approach to curating it.
What counts as memory here
The first thing to drop is the idea that memory is one thing. Inside a single project folder, several different kinds of memory run at once, each doing a job that a human team would normally split across marketing ops, sales ops, content, creative, campaigns, and the collective recall of everyone who has worked on the project.
The largest and least glamorous layer is the corpus, the raw record of the work. Meeting notes, customer transcripts, case studies, whitepapers, and blogs, along with the competitive research, email threads, and Slack discussions. My consulting repository contains thousands of files. This is the institutional knowledge that takes a lifetime to learn, and Claude reads it in seconds, treating it as context for whatever task I ask it to do.
On top of the corpus sit the instructions for using it. Skills are codified workflows, the memory of how a job gets done, and I have 43 of them. These include riveting workflows like:
Turn a podcast recording into a published episode
Cut podcast snippets into published YouTube clips
Build a competitive battlecard
Sync website and YouTube metrics into a tracker
Write a client business proposal
A CLAUDE.md file holds the standing rules for each project, the things I would otherwise repeat in every conversation. A one-page napkin file works as the running notebook, the place where lessons from one session get written down so the next session starts ahead of it.
Two more kinds of memory augment the knowledge repository and workflows. The curated memory facts, the roughly 197 entries I mentioned, act as a long-term index. These are the standing statements about how I work and where each project stands. Most are mundane on their own:
The podcast publishes every other Tuesday
Lead a LinkedIn post with the tinytechguides.com link and put Substack second
Pull-quote attribution reads David Sweenor, Founder and CEO, TinyTechGuides
Keep em dashes and colons to a minimum in everything I write
Link to the book pages on tinytechguides.com, never to Amazon
The MCP connections are the system’s memory of where everything lives outside the repository, so it reaches the current numbers instead of a stale copy I pasted in once:
Google Analytics for traffic and referral data
Google Sheets for the content calendar and trackers
Google Drive for source files and transcripts
WordPress for publishing to the blog
YouTube for podcast and channel stats
Each of these is a different kind of memory doing a different job. Remove any one of the cogs on the proverbial machine, and the works get gummed up. Together, they’re your expert sidekick.
How the memory grew
The first memories I saved were a bit ho-hum. How to read my calendar, how to format a Google Doc, where the YouTube connection lived. Reference notes, the kind of thing you scribble on a sticky note in your first week on the job. Over the months, the types of info I saved changed. The largest category in my TinyTechGuides memory today is feedback, 34 of 79 entries, and almost all of them are corrections. I told it to quit naming competitors in my research and to make sure my recommendations tie to the client value prop, not generic DIY advice. Every correction I save is a specific behavior the system then keeps me from repeating. Most recently, it started tracking live project state, which deals are open, which blog posts have shipped, and what’s up next.
I see the same thing in client work. One active engagement, call it Client A, started as an empty folder six months ago and now holds close to 500 files, among them 89 customer case studies, a 64-file design system, 51 pages of site research, 34 blog posts, 27 of its own skills, and the email and Slack threads from the relationship. When I draft or design something new for that account, I do not re-paste the brand voice or the visual rules. I point Claude to the design system, the case studies, and the earlier posts it has already read, so the context is already there.
Each client lives in its own repository, which makes every account a separate brain by design and means the memory does not fade when the work slows down, nor does it leak across clients. Another engagement, Client B, wound down months ago. The repository is still intact, with 22 meeting notes, 33 customer interview transcripts, and the email and Slack threads that captured what the engagement was trying to do. When that client comes back, or when I need to remember how I handled a particular problem, the institutional memory is sitting exactly where I left it. A folder per client gives me a memory that outlasts the engagement.
Why it compounds
The payoff from all of this is your superpower. I stopped repeating myself. Six months ago, a typical task started with five minutes of context loading, reminding the assistant who the client was, what we had agreed, and how I like the writing to sound. Today, that preamble is gone. The context is already in the folder, so the work starts there.
Skills are the clearest example. A skill is a workflow I have written down once, and once it exists, I never re-explain that job again. Better yet, I no longer write most of them for a single project. Of the 43 skills available in my TinyTechGuides repository, 25 are drawn from a common toolkit that every project uses, and only 18 are specific to TinyTechGuides. When I build a competitive brief workflow for one client, every other engagement can use the workflow as well. Now, there’s nothing specific in the skills themselves – they’re simply codified workflows on how I approach a problem – with or without AI. Essentially, it’s my codified experience.
Here’s an example of one superpower. Early on, I had to upload files to Otter for transcription. And because I’m lazy and don’t like to wait for the upload so I can hit “transcribe”, I decided to have Claude help me write a transcription service. Then, a single podcast episode took about 84 minutes to transcribe on my own machine. That was annoying, so I worked out a faster configuration using the Mac GPU, saved the recipe, and every episode since has transcribed in about 13 minutes. I solved that problem one time. Now, that memory can be used and referenced in perpetuity.
That is what compounding means in practice. I never pay twice for the same lesson, the same context, or the same setup, and the work gets faster because memory continually learns and evolves.
What I had to delete
Optimizing the system is not only about deciding what to keep, but also what to purge. A memory you never prune does not stay neutral. It rots, and then it lies to you. An outdated fact is worse than no fact at all because the assistant trusts it, acts on it, and keeps bringing it up, over and over again.
For a while, I kept an inventory of my published articles as a saved markdown (MD) file in my repo. It became outdated almost immediately, because I publish most weeks and rarely remember to update it. On June 17, I deleted it and pointed the system at the live Google Sheet instead. The flat spreadsheet I once used to track my sales pipeline met the same end after it grew into a real database. Both started as useful memory and decayed into something I had to actively work around.
So curation runs in two directions. One is capture, and I have a skill called /reflect that handles it. At the end of a work session, it asks what I learned, what was worth keeping, and what belongs in long-term memory rather than the one-page napkin. The other direction is the prune, the unglamorous habit of deleting what has gone stale or what the repository already records elsewhere more effectively.
My rule of thumb now is straightforward. Save what is lasting, non-obvious, and reusable. Delete anything that will be out of date next month, and anything the files already say. A memory system remains an asset only as long as you are willing to give it a little TLC and throw parts of it away.
Where to start
If you want to try this, don’t boil the ocean. The memory I have now took six months to grow, and most of it arrived in the form of one correction and one saved file at a time. Start by asking the assistant to remember the handful of things you re-explain every week, the way you like a draft to sound, who the audience is, and which links go first. Then put your real work in the folder, the meeting notes, the transcripts, and the case studies you already have sitting in email and shared drives. That corpus is the part that makes the assistant useful rather than generically capable. From there, let it accumulate, and prune the moment something goes stale. The system gets smarter as a byproduct of doing the work. You do not have to stop everything to build it.
The memory is the asset
A year ago, I would have said the advantage in all this was the model, whichever assistant happened to be smartest that quarter. I no longer believe that is where the edge lives. Models change every few months, and anyone can rent the same one I use. No one can rent six months of corrections, transcripts, and case studies, or the skills that know my accounts as well as I do. That took time to build, and it keeps working whether I am in the room or not. The model reads the memory, but the memory is mine. If you are going to run marketing with your trusty sidekick, treat everything it accumulates as the asset it is, and curate it like the rest of the business depends on it. A year from now, no competitor will be able to clone it.
Frequently asked questions
What is Claude memory, and how does it work for marketing?
Claude memory is the information an AI assistant keeps across sessions so it does not start every task from zero. In a marketing operation it has two parts. The first is a small set of curated facts the assistant saves about how you work. The second, and far larger, is the body of files in your project folder, the meeting notes, transcripts, case studies, and skills the assistant reads as context. Together they let the assistant produce work that reflects your accounts instead of generic output.
Is the Claude memory feature the same as the files in my project folder?
No. The memory feature stores a short list of curated facts, the kind of standing notes you would otherwise repeat in every conversation. The files in your project folder are the larger memory: the meeting notes, transcripts, case studies, blogs, and skills the assistant reads to understand the work. In my own setup the curated facts number about 197, while the project files run into the thousands. The facts act as an index, and the files do the real work.
What should I store in a Claude project to make it useful?
Store the real artifacts of the work. Meeting notes, customer interview transcripts, case studies, competitive research, prior blog posts, your brand and design rules, and the email and chat threads that explain decisions. This corpus is what makes the assistant useful on your specific accounts instead of generically capable. Add codified workflows, called skills, for the jobs you repeat, and connect live data sources so the assistant reaches current numbers rather than a stale copy you pasted in once.
How is a CLAUDE.md file different from saved memory and skills?
A CLAUDE.md file holds the standing rules for a project, the instructions you would otherwise repeat in every conversation. Saved memory is a list of curated facts about how you work and where each project stands. Skills are codified workflows that capture how a specific job gets done, from writing a blog to building a competitive battlecard. The three work together. Rules set the guardrails, facts supply the index, and skills carry the procedures across every project that shares them.
Where should I start with Claude memory for marketing?
Start small rather than building a whole system at once. Ask the assistant to remember the handful of things you re-explain every week, such as how you like a draft to sound and which links go first. Then put your real files in the project folder, the meeting notes, transcripts, and case studies you already have. Let the memory accumulate as a byproduct of doing the work, and prune anything that goes stale. The system gets smarter without a separate build project.
How does Claude memory create a competitive advantage in marketing?
The advantage is not the model, since anyone can use the same one. The advantage is the accumulated memory: months of corrections, transcripts, case studies, and skills that know your accounts as well as you do. That body of work takes time to build and cannot be copied or rented. The model reads the memory, but the memory is what you own. Treated as an asset and curated over time, it becomes an operation a competitor cannot clone.
About David Sweenor
David Sweenor is the founder and host of the Data Faces podcast, where he talks with the people who are making data, analytics, AI, and marketing work in the real world. He is also the founder of TinyTechGuides and a recognized top 25 analytics thought leader and international speaker who specializes in practical business applications of artificial intelligence and advanced analytics.
With over 25 years of hands-on experience implementing AI and analytics solutions, David has supported organizations including Alation, Alteryx, TIBCO, SAS, IBM, Dell, and Quest. His work spans marketing leadership, analytics implementation, and specialized expertise in AI, machine learning, data science, IoT, and business intelligence. David holds several patents and consistently delivers insights that bridge technical capabilities with business value.
Books
Artificial Intelligence: An Executive Guide to Make AI Work for Your Business
Generative AI Business Applications: An Executive Guide with Real-Life Examples and Case Studies
The Generative AI Practitioner’s Guide: How to Apply LLM Patterns for Enterprise Applications
The CIO’s Guide to Adopting Generative AI: Five Keys to Success
Modern B2B Marketing: A Practitioner’s Guide to Marketing Excellence
The PMM’s Prompt Playbook: Mastering Generative AI for B2B Marketing Success
Follow David on Twitter @DavidSweenor and connect with him on LinkedIn.