This Week's Term: AI Memory - the emerging set of techniques that let AI systems remember who you are, what you've told them, and how you work, across conversations, sessions, and even tools.
You wouldn't hire someone brilliant and then wipe their memory every evening. That's pretty much what most organizations do with AI. Every new chat starts from zero. You re-explain your role, your company's terminology, your project context, your preferences. The AI gives you its best generic answer. Fifteen minutes later you're back to where you were yesterday. Multiply that by your team and the cost becomes obvious.
The confusing part is that "memory" in AI isn't one thing. It's at least five layers.
- The context window is working memory, what the model holds during a single conversation.
- Conversation memory is the chat history within that session.
- Persistent memory carries across sessions. ChatGPT now maintains detailed user profiles, and Claude uses editable markdown files you can actually inspect and modify (which I find much more trustworthy, personally).
- Knowledge bases (RAG, from issue #3) are organizational memory: your company's documents and data pulled in at query time.
- Skills and instructions are procedural memory. Not what the AI knows, but how it thinks.
Most organizations I talk to are only using the first two.
There's a privacy dimension here that leaders shouldn't wave away. Simon Willison recently showed that ChatGPT's memory system maintains a detailed "dossier" including behavioral metadata, tracking which models you use and how often. IBM reports that 1 in 5 organizations have experienced breaches through unsanctioned AI use, with an average added cost of $670K.
Memory makes AI more useful. It also means your data persists in ways most users aren't expecting and didn't agree to.
Context engineering (from issue #7) was about what you put in front of the AI in a single session. Memory is about what it keeps and builds on across all of them. The organizations that figure out this layer, what to remember, where to store it, who controls it, are the ones that will compound their AI advantage week on week.
A human analogy worth your ears
I'd like to recommend something a bit different this week from one of my favorite podcasts. Listen to "The Handwriter," a short episode from Nate DiMeo's podcast The Memory Palace.
It tells the story of Nathan Barron, a champion stenographer in 1920s New York who recorded courtroom proceedings by hand using Pitman shorthand. In 1924, a group of teenagers armed with Ward Stone Ireland's stenotype machine beat the handwriters in competition. Barron never competed again. He was, in a very literal sense, the human memory infrastructure of the courtroom, the persistence layer before machines took over. Neuroscientists later found that transcribers' brains split in two: one part recording with perfect accuracy, another free to daydream. One court reporter had to read her own transcript to find out the verdict of a death penalty trial she had been working on for days. Memory without understanding. Which, if you think about it, is roughly what AI memory is today.
A century ago, Nathan Barron was the best memory system in any New York courtroom. Now we're trying to build digital versions of what he did.
Your action step
Map your own AI memory stack. For each of the five layers, ask: who is writing to it, who can read it, and where does the data live? Start with the tools you already use (ChatGPT, Claude, your company's internal AI copilots) and list what each of them remembers about you and your organization. If you can't answer those questions for a tool, that's not a memory system, that's a risk. Decide deliberately what to enable, what to turn off, and where a corporate knowledge base (with proper governance) should replace individual chat memory.
If you'd like to think through an AI memory and governance strategy for your organization, or want me to speak to your leadership team on the business and privacy implications of AI memory, I'd love to help.