TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 15—17 · FACTUAL · DURABLE
Weights, context, retrieval, memory
Only the weights are the model, and the other three are added around it.
The idea
When a chatbot seems to know something, that knowledge came from one of four places. The weights: what the model learned in training, fixed before you arrived. The context: what sits in front of it right now — your message, the file you attached, the chat so far. Retrieval: what the product fetched from outside for this answer, such as a web search or a document lookup. Memory: notes the product stored about you from earlier chats. Only the first one is the model. The other three are added around it, and they differ from product to product.
Why it matters
"The AI knows my project" is four different claims. Which place the knowledge sits in decides whether it will still be there tomorrow, whether you can check it, and whether you can delete it.
See it in the studio
You describe your site on Monday. On Tuesday the tool remembers the plot is in Hubballi. Nothing changed in the model. Either the product saved a memory note, or Monday's chat is still in the context. Test it: open a fresh chat and ask. If the site is gone, it was context. If it is still there, it was memory.
Watch for this
Believing a fact is "in the AI" when it was in your own upload. The model read your brief in context and gave it back in clean sentences. That is your knowledge, returned to you. It was never learned.
Try it
Ask a chatbot three questions: one about a famous building such as the Taj Mahal (that is in the weights), one about today's weather in your city (it needs retrieval, or it will guess), and one about something you told it last week (it needs memory). For each answer, name the place it came from. Then open the product's settings and find the memory page, if it has one.
Prove it
Name the four places an AI system can know something from, and give one example of each from a tool you use.
How it works
The weights are the trained model: millions or billions of numbers set during training and then frozen. Your conversation never changes them. The context is what the model reads on every turn — the context window is its size, and CONTEXT WINDOW on this map is about that budget. Retrieval is a product step that runs before the model answers: a search, a database lookup, a file read, with the results pasted into the context. The research paper that joined a fixed model to a live document store called the two halves parametric and non-parametric memory, which is the weights and the retrieved text in plainer words. Product memory is different again: both ChatGPT and Claude describe it as stored notes drawn from your past chats, given to the model along with your new message, and both say you can view it, edit it, turn it off or delete it. So memory is a feature of the app, not of the model. The reason this matters for a designer: a model's habits come from its weights, and a grounded answer comes from context or retrieval. THE DIET, CLOSED WORLD and RAG on this map each take one of these places further.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.