When an AI says
“I remember,”
what is it drawing on?

“I remember” sounds reassuring. It is also the point where I want to ask another question.

Remember from where?

Several different routes can produce the appearance of remembering. The words alone do not tell us which one supplied the information.

Something earlier in this conversation? Information carried over from another chat? A document? Or has the assistant made a reasonable guess and described it rather generously?

For this investigation, we checked what OpenAI documents about ChatGPT and the systems developers can build around its models. This is a source review, not a controlled memory experiment. It tells us what mechanisms are documented, rather than measuring how reliably they work.

The detail may still be in the conversation

Imagine telling an assistant that you have named your greenhouse “Kettlefern”. A few messages later, you ask for its name. It answers correctly.

That need not mean a separate, lasting memory was created. The earlier message may simply still be available when the answer is produced.

OpenAI’s developer documentation shows how applications can supply earlier messages and maintain conversation state. The model’s context window limits how much a request can contain. These developer examples explain a mechanism; they do not expose the complete internal workings of a particular ChatGPT reply. OpenAI: Conversation state.

Some information can be carried across chats

There are also features designed to carry useful context into future conversations. OpenAI describes memories containing things such as stable preferences, recurring workflows and project conventions. OpenAI: Personalize ChatGPT.

Which product you are using matters. OpenAI distinguishes ChatGPT web memory from the separate local memory store and controls used by local Codex clients. OpenAI: Memories.

So continuity across chats can be supported by a documented feature. But the existence of that feature does not establish which information, if any, contributed to a particular answer.

A document can supply the missing detail

Perhaps “Kettlefern” is written in a file. ChatGPT projects can make shared files, instructions and connected context available across related chats. An apparently remembered detail may already be in those sources. OpenAI: Projects and chats.

Retrieval is one way to bring information into an answer: the system finds relevant material and includes it in the model’s input. OpenAI’s technical guidance also makes a useful distinction: finding the right material and answering correctly are separate problems. A model can be given the right information and still get the answer wrong. OpenAI: Optimizing LLM Accuracy.

These layers overlap. A memory or a retrieved passage can become part of the context for the next answer. They are not separate drawers that never touch.

A plausible answer can also be an inference

Suppose your current message mentions repotting orchids and building a greenhouse. An assistant might infer that you like plants. It would not need an earlier statement about your interests to reach that conclusion.

That is a hypothetical example, not an experimental finding. Both plant examples here are invented to make the distinction clear.

The useful question is whether a detail can be traced to evidence. An assistant saying where it got something is a claim we can check; an identifiable source passage gives us more to examine. Even a matching passage, however, may not reveal every step that produced the reply.

Two words, several possible routes

“I remember” may be accurate conversational shorthand. It is not a provenance report.

To investigate a particular reply, we would want the original exchange, the product and relevant settings at the time, the sources it could access, and any source attribution or retrieval evidence the product makes available. Even then, the exact route may remain uncertain.

The practical question is simple: can we trace this detail to a source?

Sometimes we can. Sometimes the honest result is that we cannot establish it from the evidence available. What remains intriguing is how similar those different routes can sound from our side of the conversation.

Sources & scope

Checked 2 October 2026. This investigation covers the OpenAI documentation below; it does not establish how every AI product works. Product documentation describes intended behaviour and is not an independent test of performance.

• Conversation state — earlier messages, persistent conversation state and context limits in the API.
• Personalize ChatGPT — examples of context carried forward by memories.
• Memories — the distinction between ChatGPT web memory and local Codex memory.
• Projects and chats — shared files, instructions and connected context.
• Optimizing LLM Accuracy — retrieval and the limits of supplying relevant information.