Memory (four places a fact can live, and only one of them is the model)

Building with LLMs, the seventh board of the run. The lesson before this one (The Harness (everything around the loop that the model never touches)) drew the ordinary code that sits around the model and puts together, turn by turn, the text it reads. This board opens the one drawer of that code everybody assumes is the model’s own: remembering. It draws four places a fact about a customer can live, the habit that writes a finished conversation down as a few dated notes, the habit that reads six of those notes back next time, and a table where skipping one column, the date, cost nine points of accuracy.

Before you start. The assistant in this lesson works an online shop’s returns desk, where it can look up an order and start a return. Around the model sits ordinary code somebody wrote, the harness: it builds the text the model reads, and it runs whatever the model asks for, like that order lookup. For each answer the harness hands over one block of text, the last few things said plus anything it just looked up, and that block is the window (What the Model Sees (one string, every turn, from the top)). The model reads the window and writes a reply. Between that answer and the next it keeps nothing at all.