You Only Feel the Leash When You Leave

Every morning I get ready to go for my run, and I strap on my COROS running watch. It’s a good, light watch. When I’m ready for a new running watch, I’ll probably buy another COROS. But it would be a lie to say the decision stems only from how much I like my current watch.

Over the past couple of years, my running history has accumulated inside the COROS ecosystem. Those records are used by COROS to give me race time estimates, fitness estimates and personalised heart-rate and pace zones, recovery tracking, trends in workout intensity and volume. It’s a whole thing. I bought a watch and acquired an ongoing relationship with its opinions about my fitness. Having all of that in one place is valuable to me.

I can export my activities from COROS. The files could follow me to another service. But I can’t be sure how much of the experience I value would travel with those files to another provider. The uncertainty alone works in favour of staying with what I know, and with what knows me.

When it comes to choosing an AI services provider, people often explain their choice with reference to the capabilities of the models, what their employer makes available, or their familiarity with the tools. It is much less common for me to hear that people stay because the tools know them.

I think that this will change. As useful context accumulates with one provider, keeping that context becomes part of your choice to stay with that provider. Other models may perform better on your tasks, but to get their full benefit, you would first have to provide the background, correct the assumptions, and restate the earlier decisions. At a certain point, the difference in performance stops being worth the effort it takes to move.

I didn’t think about any of this when I bought my running watch. I was in the ecosystem before I realised I was. With AI, I want the effort I spend explaining things to stay useful when I change providers. The history and memory should be something I can carry into that next choice.

Two different AI systems from two different companies helped me prepare this essay: Claude, from Anthropic, and Codex, from OpenAI. Both of them could assist me inside the same drafting tool that I created specifically for this purpose, and both could read the same memories. The record of memories lives in a separate system that I control. I choose which agents can read from it and add to it.

I call my personal memory vault my “Memory Palace”, named after the Method of Loci technique for information recall. It holds a continuing account of my life and projects, and how they fit together. Keeping that account in a system that I control was the main reason I built it. I want to be able to change my mind about my AI providers, and to move between models freely without re-explaining myself or leaving context behind. The Memory Palace lets me do that.

During one review of this essay, Codex (GPT-6 Astra) offered an opinion about the impression it might leave on professional colleagues vs. an audience of more casual readers. I hadn’t asked for that assessment this time. It had retrieved a concern I’d raised during an earlier Obiter review, because these essays are read both by the people I work with, and also a wider audience of people I don’t personally know. I care how these essays land, and it was useful not to have to explain that concern again. I want that sort of continuity in every tool I use.

The value of an AI tool increases greatly as it is able to learn what matters to you. In June, OpenAI updated ChatGPT’s memory system, saying that memory is “central to making ChatGPT more useful: knowing you, helping you, and doing more for you.” I agree completely, and that is reflected in the time I’ve spent building my own memory system.

Memory also changes what we’re choosing when we’re picking a provider. Models change frequently, but memory accumulates between releases. I feel strongly that for heavy users, accumulated and well-integrated memory will come to matter more than the models.

The well-integrated point matters a lot, too. I reached a point where my memory system got good enough to cause problems. The agents could find everything. Getting them to work out whether any of it mattered to the conversation we were actually having was another project.

To manage this, I added a short-term memory layer with event decay to keep agents focused on what is happening for me right now. The complete record still sits underneath, and each night, a synthesis run considers how new facts fit into the threads that they belong to. An old fact can become important again because a story is continuing. A recent fact can matter very little because its story has been resolved.

Frontier providers are working on this same puzzle. OpenAI has described a background process that keeps memory current, including recognising when an event moves from upcoming to past. Accumulated history is most helpful when the system can determine which parts continue to matter.

Some providers already make explicit commitments to portability. Anthropic says that all memory data is included in its exports and offers tools for moving memory between services. Asking AI to prepare a handover can reduce the effort of moving, but the memory import features are described by Anthropic as being experimental, and so successful incorporation of the supplied data is not guaranteed.

The ability to export memory does not, by itself, put the problem to rest. It is important to me that a memory export includes the complete account the incumbent system built about me, including the corrections, decisions and clarifications made along the way. Those distinctions are a big part of what I’ve spent my time teaching the system. An export option alone doesn’t ensure that they are all retained in the data I can take with me.

My Memory Palace doesn’t guarantee that different models will interpret the same history in the same way. What I control is the account that they work from. A nightly synthesis produces accounts saved in files that I control, while also integrating the underlying records. I can change the agent doing that work without losing what they did. Models from individual providers still process my session context, but I don’t depend on the provider I’m leaving to supply my continuing account when I move my memories from one model to another.

Dependence can develop even where export functionality remains available. If leaving requires someone to rebuild some of the understanding held by an AI assistant, the cost of switching is already there. As the service gets better, that cost can grow.

My larger concern, and the main reason I maintain my own memory system, is that providers may eventually make moving harder by restricting access to memory. My prediction is that this becomes more likely if one or a few frontier providers can establish durable market dominance. Making it easy to move helps a challenger win customers. Dominant providers have more existing relationships to lose. I am not comfortable with my ability to leave being dependent on the incentive remaining favourable.

As a lawyer, I would like laws and regulations to preserve these choices. But the existing rights don’t map cleanly onto the things that make memory useful. The EU GDPR includes a qualified right to move personal data supplied by a user. The portability guidance endorsed by the European Data Protection Board includes observed data (such as a record of how a service is used), but generally excludes inferences and derived data created by the provider. A memory may combine something I expressly told the system with a conclusion it drew from several conversations. Those elements may attract different rights. An inference may still be accessible under a separate right of access, even when the portability right doesn’t cover it.

Australia also maintains access rights under the Privacy Act, but there is no general right to interoperable transfer from one provider to another. The Consumer Data Right provides sector-specific data sharing requirements in banking and energy, with a staged extension to non-bank lenders. It doesn’t provide an equivalent portability regime for AI services. These rights provide precedents for protections I would like to see extended to AI services, but they don’t currently establish a right to take the full account from one AI system to another.

My own memory system is the product of hundreds of hours and requires daily maintenance. Apparently “I’d like to keep my options open” needed to become a second job. I’m willing to do it, but I would quite like this to become less of an undertaking. I would like keeping a usable account of our own history, independent of whichever AI provider we choose, to become an ordinary part of using these tools. For now, I’ve chosen to do the work myself.

My next running watch will probably be another COROS. The accumulated value of my running history makes staying feel easier than finding out what would follow me elsewhere.

The account my AI tools build reaches across my work, projects and personal decisions, and I want more freedom over who I rely on to maintain it.

With AI, I’ve noticed the memory being built, and this time I’ve decided that I’m going to be the one who holds it.