Chatbot: conversation lives outside the weights

A chatbot is a conversational application. Its interface may look like a dialogue with one persistent participant, while each model call processes a constructed input containing the history the application chooses to include.

A language model can be trained for assistant behaviour, but the surrounding application manages messages, tools and any longer-term memory.

Three different kinds of memory

Model weights hold learned patterns from training. Conversation context holds the messages available to the current computation. A retrieval store can hold information selected for later insertion into that context.

Deleting a message from the next request changes the available context, not the model weights. Saving a conversation in a database does not automatically make its contents visible to the model either.

As the dialogue grows, the application may retain the full history, summarize it or retrieve selected parts. These choices change continuity, latency and resource use. The chat interface hides that context-management work, but does not remove it.