Chapter 7 — Separating Context, Memory, and Training Data

"What the AI knows" is not one single thing

When using AI, it can feel like:

• information conveyed in this conversation right now
• information discussed in a previous conversation
• saved information
• information the AI used for training
• information retrieved through search or from files

are all part of one "AI memory." But treating these as the same thing makes it hard to judge AI's behavior correctly. This chapter separates the following four in particular:

Context
Memory / saved information
Training Data
learned model state

Context

Context is a concept related to the information available to a model or AI service for the current process. For example, in a conversation, a condition stated a little earlier can affect a later reply. But once said = always available as Context does not hold. How much information is currently available may differ depending on the service and conditions. We don't guess at the specific, undisclosed mechanism of context assembly.

"Said before" and "usable now" are different

Suppose you conveyed some information in an earlier conversation. The fact that follows from this is only "that information was discussed before." From this alone, you cannot say "it's therefore available in the current interaction." Whether it's currently available, referenced through a saved-information feature, or something else, we leave it distinguished when it cannot be confirmed.

Memory / saved information

Depending on the AI service, a feature — sometimes called Memory — may be provided that stores information about the user. Information stored by such a feature can be thought of as saved / persistent information. However, saved information = model parameter does not hold. We do not speculate beyond public information about how saved information is internally retained, or how it's used in the current interaction.

Training Data

Training Data is data used to Train a model. In Training, the data is used to adjust the model's parameters or learned state. What matters here is the distinction Training Data != current input. The text you type into a model right now is input used in the current Inference. Whether it becomes Training Data is a separate question.

Learned model state

A trained model has parameters and a learned state formed through Training. We keep this distinct from:

• current context
• saved information
• runtime retrieval
• current input

For example, even if an AI can answer a question about some piece of general knowledge, that single output alone does not let you pin down that "this text was learned as Training Data."

Retrieved information is also separate

Chapter 8 covers Retrieval. There is a mechanism where information is obtained from external documents or files through Retrieval and used in generation. That information, too, is not automatically equated with Training Data.

Conceptually, we can separate it this way:

Training Data
    ↓
Training
    ↓
learned model state

current input ──────┐
Current Context ────┤
Saved Information ──┤
Retrieved Info ─────┤
                  ↓
              Inference
                  ↓
                output

This diagram exists to avoid confusing the roles of different kinds of information.

A single word — "I remember" — is not enough to determine anything

Suppose an AI answers "I remember." That sentence alone does not let you uniquely determine which of the following is going on:

• it remains in current context
• it is referencing saved information
• it was retrieved
• it relates to learned model state
• some other service-side mechanism is involved

The fact that this sentence was generated and which technical mechanism was actually used are two different things.

Don't assume a conversation automatically becomes Training Data

Whether information a user inputs to an AI is used in future Training depends on the service's public information and terms, which need to be checked. This book does not draw the inference: had a conversation → automatically became Training Data. Conversely, we also don't assert it will absolutely never become Training Data without Evidence. Where it cannot be confirmed, we treat it as unconfirmed.

Lining up the four

These are not a complete taxonomy of AI-related information. They are four important concepts lined up to avoid confusion in this book.

Chapter summary

Context != Memory != Training Data
Saved information != learned model state
Inference input != Training Data
Retrieved information != Training Data
an "I remember" output != proof of a memory mechanism

• Context relates to information currently available
• something said in the past and something currently available are not the same
• information saved through Memory and model parameters are not the same thing
• Training Data is data used for Training
• current input or retrieved information is not automatically treated as Training Data
• learned model state and current context are kept separate
• undisclosed internal processing such as saving, context assembly, or use in training is not filled in with guesses

The next chapter organizes how Search, Retrieval, RAG, Tools, and Agents extend a model's capability beyond the model alone.