Training data is the text the model learned from and is not stored after training — only its influence on the weights remains. Parameters are the learned weights themselves, static at inference time. Context is the active input the model reads right now, cleared at the end of the session. Memory is application-managed information saved externally and retrieved into future conversations.

These four terms get conflated constantly, but they're distinct layers. Training data is the large collection of text — web, books, code, conversations — used to train the model; after training it isn't stored anywhere retrievable, only its effect on the weights remains. Parameters are the learned weights of the network: they store patterns from training but not as a searchable database — knowledge is distributed across millions or billions of weights, and parameters don't change at inference time unless you fine-tune. Context is the current input given to the model during inference — user prompt, conversation history, system instructions, retrieved documents, tool outputs — and it's temporary, cleared at the end of the session; the model only "sees" what's in its current context window. Memory is user-specific information saved externally in a database or vector store and retrieved into future conversations; it's managed by the application, not the model itself, and is what enables personalization and recall across sessions without retraining.