Retrieval vs Memory: How LLMs Know Things About Your Brand

Memory: what the model learned in training

A language model is trained on a large snapshot of text. Whatever your brand looked like across the web at that time is roughly what the model remembers. If you were discussed often and consistently, it holds a stable picture. If you were rare or contradictory, it holds a blurry one and fills gaps by analogy to similar companies.

Memory has three properties worth understanding.

It is frozen. The training cutoff is a wall. Your new pricing, your rebrand, your new product line do not exist on the far side of it.

It is probabilistic. The model does not store facts; it stores patterns. Ask it your founding year and it produces the most likely number, which may be right, may be a competitor's, may be invented.

It is slow to change. You cannot update it. The next model version may know more, and it will take months to arrive.

Retrieval: what the system looks up

Most answer engines now retrieve before they generate. The question goes to a search layer, which returns a handful of documents, and those are placed in front of the model with the question. The model writes an answer using them and, usually, cites them.

Retrieval has the opposite properties.

It is live. A page you publish today can be in an answer this week, sometimes the same day.

It is literal. The model tends to quote or closely paraphrase what the retrieved page says. If your page says "from $49 per month," that is what appears.

It is controllable. You decide what is on the page, whether crawlers can reach it, and whether it renders without JavaScript.

How to tell which one you are seeing

A few practical signals.

If the answer has citations, it is retrieval, at least in part. If it has none and the engine normally cites, the model answered from memory.

If the answer gets a recent change right, that is retrieval. If it states old pricing or a discontinued product, that is memory.

If you ask the same question with web access off and the answer changes substantially, the difference is retrieval.

In our testing, the major engines lean on retrieval for factual questions about specific companies, and on memory for broad category questions. "What does Acme do" is retrieval. "What are the best tools for X" is often memory with retrieval layered on top, which is why the same old names keep appearing.

Why this matters for the work

The work you can do this month only affects retrieval. That is not a limitation; it is where the effort pays.

Quotable pages, crawler access, no JavaScript dependence, schema, and consistent facts across sources all change what gets retrieved and what the model does with it. The effects show up in weeks.

Memory responds too, but on the next training cycle, and only in proportion to how much of the web agrees about you. The best way to improve what the next model remembers is to improve what the current model retrieves and let the web catch up.

"Get the model to know us" is not a task. "Get the model to find and quote the right page" is.

A diagnostic you can run

Take your five most important facts: what you do, who it is for, price, location, and the one thing that distinguishes you. Ask each engine for them with web access on and off, where the interface allows it. Note which facts are right in each mode and which sources are cited.

Facts wrong in both modes mean the web itself is inconsistent; the fix is entity consistency across your site and third-party profiles. Facts wrong only without retrieval mean memory is stale and retrieval is working, which is the healthy state. Facts wrong only with retrieval mean the engine is finding the wrong page; the fix is on your site.

At AEO GEO Labs this split is part of the first audit, because it determines whether the work is mostly technical, mostly content, or mostly third-party sources. For that diagnosis across your full prompt set with the sources mapped, AEO GEO Labs delivers it as the starting point for everything that follows.