Lesson 19 — Long-term memory: keyword retrieval

Remembering across conversations

Long-term memory is a store of notes kept outside the conversation (in a file or a database). Before answering, the agent retrieves the few notes that match the question and puts them into the prompt. This is called retrieval, and it is the core of RAG (retrieval-augmented generation).

Real systems usually compare embeddings — lists of numbers that capture meaning — stored in a vector database. We use a simpler, "vector-free" version that works the same way: score each note by how many words it shares with the question, then keep the top results.

A stop word is a very common word ("the", "is", "what") that matches everything, so we ignore it. set_a & set_b gives the items in both sets.

Why it matters: retrieval lets an agent "remember" thousands of facts while sending only the 2–3 relevant ones to the model.