Two people can have an equally strong bond with AI but very different control over what they use. One manages their own files; another has access to an account in someone else’s application. A third paid once and only discovers what they bought when problems arise. How it is paid for tells us only part of its independence: a free service can be closed, and your own agent may regularly consume paid services.
The model is not the agent’s whole identity
Here, by an agent we mean a combination of model, instructions, memory, history and possibly tools for action. The model forms responses within it, but is not their only source material. OpenClaw, for example, uses separate files for a name and characteristics, intended personality and long-term notes. Model choice is another part of the settings.[1]
These materials can remain when the model is replaced. The new model builds on the same information and work in progress. The user may recognise a familiar counterpart with a different speech rhythm, humour or approach to a task. Such continuity has a technical basis; changing the model does not itself prescribe that the person must begin a new relationship.
What is preserved in practice depends on the models and which stored information actually reaches the response context. Files can stay the same while expression changes. And a memory backup need not contain history, voice, tool settings or every other part of the setup.[1:1]
Your own agent may still use a remote model through a paid interface. You then manage part of the setup while the model’s availability depends on a supplier. Running important components locally can give greater control over their version and operation. Check where data are stored and where they travel during use separately: even an agent with local memory may pass them to a remote model, voice service or search.[1:2]
Why it can know something and not remember it another time
The context window is a limited space for material the model currently works with when creating a response. It may contain your new message, part of the conversation, instructions, selected notes or text obtained by a tool. It need not contain the entire history you see on screen. Over longer use, older parts may be summarised or only selected information loaded.[1:3]
Long-term memory stores information for further use. But storage does not yet say a particular note reached this response. An assistant may have a file with your dog’s name available but not use it in a particular conversation. Or it loads an old summary missing an important correction. Selection and management differ between services.[2][1:4]
Your experience of continuity can therefore have a real basis while also being imperfect. When something important drops out, first try naming what you need to bring back: ‘Last time we agreed on this. Here is that section.’ In memory management, then check whether the information is stored and current. One forgotten sentence does not yet tell you whether data were lost, the model changed or simply the right material was not loaded.
The assistant’s assumptions can enter memory too. If a summary presents its older interpretation as your own view, correct the specific record. You do not have to erase the whole shared history. You are correcting information that began steering another conversation elsewhere.
A general service and your own agent are different arrangements
With an ordinary account, you change selected settings; the operator determines the model offering and rules. Personally meaningful contact may nevertheless develop in a product not designed primarily for relationship continuity. A subscription alone does not guarantee an unchanging assistant.
A specific historical example is GPT-4o’s retirement from ordinary ChatGPT on 13 February 2026, with old conversations continuing on replacement models. This announcement did not itself mean GPT-4o was simultaneously removed from the API.[3] A change may be driven by technical development, safety or product direction. Its effect on a person may be considerable even without an intention to sell more intimacy.
The practical question is: what can you export, and what can you restore elsewhere from it? Sometimes you gain a path to continuation; sometimes only a message archive without the voice, memory or other parts you value. An administrator or support team can help distinguish these. You do not first have to run your own server to understand your options.
I paid once. What remains mine?
| What you acquired | What to clarify |
|---|---|
| Software and model files for local use | What they need to run, whether they require someone else’s server and how to restore them. |
| Building your own agent | Who manages accounts, memory and access, and who continues paying for the model or hosting. |
| Long-term or ‘lifetime’ cloud access | Exactly what the offer covers and what happens when the service changes or ends. |
| A credit bundle | What consumes credits, when they expire and what remains when they run out. |
The price on the receipt is only one part of the conditions. A one-off payment may still require someone else’s infrastructure. A regular cost may instead pay for a remote model’s consumption by your agent or a clear flat rate without further incentives. What is actually provided for the price is what matters.
Card G helps you record where your AI’s identity, memory, model and other components are. An empty field is a question for the administrator, not a poor mark in technical knowledge. Passwords and keys do not belong on a shared card.
Sources and notes for this chapter
OpenClaw. Agent runtime, Memory overview, Models CLI, Model providers. Agent, memory, models, providers. Documentation loaded 11 September 2026. Establishes the separation of configuration and memory material from model selection. Conclusions about possible continuity of contact are a technical interpretation of this arrangement, not a measurement of identical experience or a guarantee of the same behaviour. ↩︎ ↩︎ ↩︎ ↩︎ ↩︎
OpenAI. Memory FAQ: help.openai.com/en/articles/8590148-memory-faq. State reviewed 10 September 2026. Documentation of memory and its management, not a study of clinical understanding of a user. ↩︎
OpenAI. Retiring GPT-4o and other ChatGPT models. Official announcement. Reviewed 11 September 2026. A historical example of retirement from ChatGPT on 13 February 2026 and the distinction from the API, not a claim about today’s availability of all the models mentioned. ↩︎