I use several AI tools, often for different purposes. I might work through an idea in ChatGPT, develop something further in Claude, or use Code to work on a software project. Which tool I use depends on what I’m trying to accomplish and which one is best suited to the task at the time. AI is advancing so quickly that the best tool for a particular task can change in a matter of months or even weeks. Companies are constantly leapfrogging one another, and I expect that to continue.
The problem is that the information I develop in one conversation doesn’t reliably carry over to the next, even within the same AI tool, much less when I switch between platforms.
I might spend hours discussing a project, exploring different approaches, explaining my reasoning, and making decisions. Over time, the AI develops a useful understanding of what I’m working on, what I’ve considered, and where I want to go next.
But as conversations grow longer, the thread becomes less reliable and accurate. It may lose track of earlier details, overlook decisions we’ve already made, or contradict something we previously established. Eventually, starting a new conversation becomes the better option.
And that’s where the problem begins.
I end up explaining the same things again: what the project is about, what I’ve already tried, why I made certain decisions, and what I’m still considering.
Switching to another platform creates the same problem. I can use whichever tool is suited best for the task, but first I have to bring it up to speed. Keeping everything within one platform would avoid some of that inconvenience, but then I’d be tying years of accumulated knowledge to a tool that may no longer be the best choice six months from now.
This matters more as AI becomes a larger part of my work and everyday life. I’m using it to develop ideas, solve problems, manage projects, and explore my own thinking.
Through all of this, I’m accumulating knowledge about my projects, interests, priorities, and the reasoning behind my decisions. Much of that knowledge becomes difficult to carry forward when I start another conversation or change tools.
That led me to a question: Why am I allowing an AI tool to determine where all this accumulated knowledge lives?
I’ve been treating each conversation as a place to do work, but those conversations also create something valuable. They contain research, decisions, explanations, ideas, and the reasoning behind much of what I’m doing. Some of that information is already captured in documents and project files, but a great deal of the context exists only in the conversations and the projects themselves.
The problem isn’t that the information is gone. I can return to an old conversation and find what I need. But that means remembering where we discussed something, searching through old threads, or manually reconstructing the context for a new one.
I want my accumulated knowledge stored in one place, accessible to any AI tool rather than tied to a particular platform.
Instead of keeping separate histories scattered across ChatGPT, Claude, and other tools, I want a place to preserve the information that matters: what I’m working on, what I’ve learned, which decisions I’ve made, why I’ve made them, and what I intend to do next.
Then, whichever tool I choose to use can draw from that same history.
The goal is to start a new conversation without rebuilding the context from scratch. If another company releases a better tool, I want to use it without leaving behind years of accumulated work. And when I return to a project after several months, I want to pick up with an understanding of where things stand and how I’ve gotten there.
The idea isn’t to make every AI remember every conversation I’ve ever had. Much of that information wouldn’t be useful anyway.
I want to preserve the knowledge that remains valuable after the conversation ends.
And I want that knowledge to be mine to organize, maintain, and carry forward, regardless of which tools I use to access it.
There’s another reason I want to preserve this information.
I’ve come to understand that AI is exceptionally good at working with large amounts of information, finding connections, and identifying patterns. But it can only analyze the information it has access to.
The accumulated knowledge may also be valuable for reasons beyond simply remembering what I’ve already done.
What might become possible if I gave it access to years of my experiences, decisions, conversations, ideas, and observations?
I’m not sure what it will reveal, but I’m interested in finding out.
That gives me another reason to capture the information now, rather than waiting until I know exactly what I can do with it.
Of course, I’m not the first person to recognize the value of having a central place to preserve and organize knowledge. Applications like Obsidian are built around that idea, allowing people to maintain personal knowledge bases and connect information across notes, projects, and ideas. Other products, including Mem0 and Supermemory, focus more specifically on giving AI tools access to shared memory.
I could have used these existing solutions rather than building something myself.
But I wanted a system organized around my particular needs, one that can integrate with the tools I already use and evolve as those needs change. I’m not simply looking for a better place to store notes or a way for AI to remember my preferences. I want a shared knowledge system that different AI tools can actively use and help maintain as I work.
So I started building my own.
The idea is fairly straightforward.
My original documents, voice notes, research, and other source material live in Google Drive. A separate database organizes useful information from those sources, including projects, decisions, ideas, questions, and the relationships between them. Each AI tool has custom instructions that explain how to access and work with this shared knowledge, while individual projects within those tools have additional instructions specific to their purpose. Together, these instructions guide the AI in retrieving relevant context, recording new information, updating project progress, and maintaining continuity across conversations and platforms.
The goal isn’t to make every conversation part of one enormous, permanent chat. It’s to preserve the useful knowledge that comes out of those conversations so I can return to it later, regardless of which AI tool I’m using.
My writing process is a good example.
I started making voice notes by dictating my thoughts directly into Google Docs. Over time, that evolved into conversations with AI using voice mode. Rather than simply recording my thoughts, I began using these conversations to work through problems, explore ideas, discuss projects, and think through decisions. Sometimes we discuss ordinary events from my day. Other times, much of the conversation explores a particular question, experience, or problem I’m trying to solve. These conversations then become voice notes in my Knowledge System, preserving not only what I’m thinking about but also how I work through it.
The system helps organize that information so it doesn’t remain buried in individual conversations. For example, ideas that might be worth developing into articles are captured in a writing inventory. When I’m ready to explore one further, the relevant notes and previous discussions can be gathered into a writing brief that identifies the central question, supporting experiences, and possible direction for the essay.
That’s how this article began.
The system continues to evolve as I discover new ways to use it. What began as a way to preserve information is expanding into a system that helps me manage projects, track tasks, and decide where to focus my attention.
For example, I’ve also developed a dashboard that brings together tasks across all of my projects and AI tools. It identifies tasks I can work through in voice mode while I’m walking or running, others that AI can handle independently, and those that require my direct attention. The goal is not simply to keep track of what needs to be done, but to help me identify the most valuable thing to work on next.
In addition to a memory system, it’s becoming a way to put that memory to work.
Of course, building a system like this comes with its own complications.
Recently, while reviewing whether everything was working as intended, I discovered that my knowledge was actually spread across two separate databases. One contained information that hadn’t made it into the other.
There’s some irony in that. I’d built a system intended to eliminate fragmented knowledge, only to discover that the system itself had fragmented.
Having a shared database also doesn’t necessarily mean the AI tools will use the information correctly.
In a separate test, Claude successfully connected to the database but hadn’t retrieved an updated work queue that was already saved there. The information was available, yet the assistant hadn’t incorporated it into its work.
Giving AI access to information is one thing. Getting it to consistently retrieve, use, and update that information is another.
And keeping that information up to date introduces another challenge. Ideas evolve, decisions change, and priorities shift. The system needs to recognize those changes rather than continue relying on information that may no longer be relevant.
A system that remembers everything but can’t distinguish what’s current from what’s outdated could create as many problems as it solves.
The challenge isn’t simply collecting information or making it accessible to different AI tools. It’s making sure that information remains accurate, relevant, and consistently maintained.
Figuring out how to make these tools do what I want is an ongoing process. And what they can’t reliably do today, they may be able to do before long.
Since I started using AI, I’ve often tried to make these tools do more than they can reliably handle. I’ll have an idea, experiment with different approaches, and sometimes discover the technology simply isn’t ready. Other times, I make progress only to encounter a new limitation that requires rethinking what I’ve already built.
It’s often a process of taking two steps back to move five steps forward.
Then, months later, a new feature comes out that makes the same task relatively easy.
Something I’ve spent hours trying to figure out eventually becomes a built-in capability.
I suspect the Knowledge System will follow a similar path.
As AI continues to develop, I expect sharing information across conversations, projects, and even competing platforms will become much easier. What I’m building myself today will most likely become standard functionality.
Which raises a reasonable question: Why spend time building something that I might eventually be able to use straight out of the box?
It’s a question I’ve considered.
I enjoy experimenting with these tools. I like exploring what’s possible, finding ways around current limitations, and discovering new ways to use the technology. Sometimes those experiments produce something useful.
In some cases, waiting would probably be more efficient. But part of the value comes from working through the problem myself. I learn what works, where things break down, and what I might try differently next time. That experience gives me something to draw on in other projects, something to write about, and a better understanding of what I’m asking these tools to do.
A future update might make the system I’m building unnecessary. It wouldn’t make the experience of building it unnecessary.
What interests me most about the Knowledge System isn’t necessarily what it can do today, but what might become possible as the information accumulates and AI continues to improve.
While working on this article, I decided to put that idea to the test. I asked the system to examine the information it had already collected and identify patterns, connections, or insights that might be useful.
Some of what it found was particularly interesting.
Examining recurring assumptions. The system connected conversations about achievement, effort, and personal boundaries and identified underlying assumptions I hadn’t explicitly spelled out in those discussions. These are beliefs I’ve already come to recognize in myself. That gives me some confidence that it might also help me recognize patterns I haven’t yet understood as clearly.
Recognizing new opportunities. The system connected several AI applications I’ve been developing for my real estate business with earlier conversations about consulting and writing about AI implementation. It suggested that the knowledge and experience I’m accumulating while building these tools might have value independently of the applications themselves. I’ve already considered some of those possibilities, but the system offers another way of looking at how these projects might fit together.
Seeing the system draw useful connections across these conversations is encouraging, especially since I’ve only been collecting this information for a short time.
A longer record will let me revisit past decisions, compare my reasoning with what actually happened, and see how my thinking develops along the way. I can look at what experience teaches me and whether those lessons influence what I do next.
But those are only the uses I can imagine today. As the record grows and AI develops, I expect this information to become useful in ways I haven’t yet considered.
That expectation is part of why I’m gathering it now. I don’t need to anticipate every future use to see value in preserving the experiences and reasoning those uses might depend on.
Related Essays
The Flat Top: where the Architecture of Identity series begins.
Actions, Not Outcomes: pursuing a result without making it the measure of myself.