I only built shared AI memory because I was scared of losing it

I don’t think I would ever have thought seriously about shared AI memory if I hadn’t run into a very ordinary account problem. I wanted to separate my work and personal AI use. That was basically it.

I had started using AI heavily while logged in with my work email, and over time that stopped being just a convenient login. The accounts had accumulated context: conversations about Editorial Intelligence, early versions of ideas, things I’d rejected, arguments I’d returned to several times, decisions about the website, experiments with workflows, explanations of what I meant when I used particular terms. None of it felt particularly important as infrastructure while I was creating it. It was just talking to ChatGPT and Claude.

Then I wanted to move. When I tried to move my OpenAI account away from my work email to a personal one, I couldn’t simply change the identity and take all of that accumulated history with me. I could create a new personal account, but that would mean leaving the existing context behind. So if I wanted continuity, I was effectively forced to keep using the work email.

That immediately exposed the future problem: at some point I will leave Sage, and when I do I will lose access to that email and the AI account attached to it. I did not want Editorial Intelligence to disappear with the account. Claude reinforced the same problem from another direction — I could not simply move the accumulated work from the work-linked account into a new personal identity there either.

That was the moment this stopped being an abstract AI-memory question for me. I realised I had built quite a lot of Editorial Intelligence inside accounts I might not always control.

The obvious problem I had somehow ignored

There is an even simpler version of this. At some point I will leave my current employer, and when that happens, work accounts won’t come with me — nor should company information. That is normal and necessary. But it made me confront a distinction I had not been managing particularly well.

My job and my independent thinking regularly intersect. An experience at work might trigger an Editorial Intelligence idea. Something I learn while researching Editorial Intelligence might change how I approach a problem at work. A new workflow might start as a personal experiment and then make me see an organisational problem differently. The intellectual boundary is not always as clean as the account boundary.

That does not mean proprietary company information should migrate into a personal system — it should not. But my own arguments, public research, methodologies, experiments and developing thinking need to survive a change of employer. They also need to survive a change of AI provider. I had not designed a system for that because I had not realised I needed one. The email problem forced me to.

This is why I started building shared memory

My first reaction was essentially: I need somewhere else for this stuff to live. Not another AI. Somewhere the AIs could both use. That eventually became GitHub. The important part was not GitHub itself — I could have chosen another durable knowledge system. What mattered was changing the architecture.

Before, the pattern was effectively:

Me → ChatGPT

and separately:

Me → Claude

Each relationship accumulated its own history. Now the pattern is closer to:

Me → durable knowledge layer ← multiple models

Claude and ChatGPT can work on the same Editorial Intelligence material. One can create something and the other can challenge it. I can return months later and recover a decision from the repository rather than hoping the right conversation is still available. If one account disappears, the project does not disappear with it. If one provider gets worse, I do not have to reconstruct months of thinking before I can switch.

That sounds obvious to me now. It absolutely was not obvious before the account problem.

The strange thing about AI lock-in

Software lock-in normally looks quite visible: your files use a proprietary format, your database is difficult to migrate, your company has spent millions integrating one system. AI lock-in feels different. It can develop through familiarity — the model knows what you’re doing, it remembers previous conversations, it understands your terminology, you don’t have to keep explaining yourself. Every one of those things makes the product better, and every one potentially increases the cost of leaving.

The valuable thing is not necessarily the model. It may be the accumulated context surrounding your relationship with the model. That is why I’ve become increasingly interested in the distinction between model intelligence and user intelligence. The model belongs to OpenAI, Anthropic, Google or whoever comes next. The accumulated understanding of my work shouldn’t.

I wrote in Context is capital that the advantage in AI-assisted work increasingly comes from the material surrounding the prompt: decisions, history, evidence, constraints and accumulated judgement. The account problem added an uncomfortable corollary. If context is capital, then concentrating it inside an account you cannot easily move is a form of dependency.

Files weren’t enough

Originally I thought this was mainly a backup problem: keep the important files somewhere else. But the more I worked on Editorial Intelligence, the more inadequate that started to feel. A finished article does not tell me everything that produced it. It does not necessarily contain the competing argument we rejected. It does not tell me which evidence made me change direction. It does not contain the question I have not answered yet. It does not explain why two apparently similar ideas are actually different.

The useful layer is often between the files. That is what eventually led me towards the idea of a save point. A useful AI save point is not a transcript of everything that happened. It is enough durable context for another model, another person or future me to understand where the work currently stands and continue.

I developed that argument properly in Your AI work needs a save point. The practical direction also appears in The workflow should outlive the model, GitHub for knowledge and Build your own portable intelligence repository.

Those pieces now look like a fairly deliberate body of thinking. They did not start that way. A lot of them trace back to something much less sophisticated: I wanted to change the identity attached to my AI use and discovered that moving the account was not the same thing as moving the accumulated intelligence.

I think this problem will get bigger

Right now, many people could probably lose an AI account and be annoyed rather than professionally damaged. That may change.

Imagine someone who has worked with an AI assistant for five years. It knows the projects they have completed, their research, how they make decisions, what they have learned from previous mistakes, which approaches repeatedly fail, their writing patterns, their unfinished ideas, their own accumulated system of work. Then they change employer. Or the company switches AI provider. Or their account is suspended. Or the product gets discontinued. Or a new model becomes dramatically better. What exactly moves with them?

We talk a lot about AI becoming a second brain. I’m not sure enough people are asking who controls that brain’s memory.

The recent discussion around people losing access to AI accounts makes the catastrophic version of this easy to understand. But I suspect the more important version will be mundane: people change jobs, they change email addresses, companies change software contracts, people decide they prefer another model. Those should not be events that force knowledge workers to reconstruct years of context.

The thing I wouldn’t do now

I would not deliberately build an important body of work so that it could only be recovered from one AI account. That does not mean avoiding ChatGPT memory or Claude projects — they are useful. It means I no longer want them to be authoritative. They can remember me. They can make the interface easier. They can recover useful context. But for work that I expect to matter in several years, I want a version of the knowledge that I can inspect, maintain and move.

The model should be able to leave. The account should be able to disappear. My employer should be able to change. The work should still have somewhere to resume from.

The funny thing is that I now see shared memory as one of the most important pieces of my AI setup. But I did not arrive there by designing an AI architecture. I arrived there because I tried to change an email address, and realised how much I might leave behind.