AI memory vs. context: what is the difference?

People use "context" and "memory" as if they meant the same thing. For an AI assistant they are different, and the difference explains a lot of confusing behavior.

Context is what the model can see right now

Context is the text a model reads when it produces a reply: the conversation so far, the instructions, any attached files. It exists only for that request. When the next request begins, the model starts from nothing and reads whatever the app sends again. For the background on that limit, see what a context window is.

Memory is what persists between conversations

Memory is information that survives after a conversation ends, so that the next one can use it. A model does not do this by itself. Memory is a feature that the app, or the way the model is built, has to provide, and there is more than one way to provide it.

The main approaches

• Resending the chat history. The simplest approach. It works until the history outgrows the context window, and it ends when you open a new chat.
• Summaries. The app condenses older conversation into a short note and includes the note instead of the full text. It is compact but lossy: a detail that did not make the summary is gone.
• Saved facts or notes. The app, or you, store short statements such as "prefers metric units" and add them to future requests. They are easy to inspect and edit, but only as complete as what was saved.
• Retrieval. Past conversations and documents are stored and searched, and the best matches are added to the context for each question. The pattern is compared with long context in RAG vs. long context. It scales to a lot of material, but answers are only as good as the search: if the right passage is not retrieved, the model never sees it.
• Changing the model itself. Training or fine-tuning can teach a model patterns and style. It is slow and expensive, and a poor fit for personal facts that change, so it is not how an assistant remembers your last conversation.

Real products usually combine several of these.

One fact, four ways to lose it

Suppose you spend a week planning a trip with an assistant, and on day one you mention that you are vegetarian and travelling with a toddler.

• With history only, that fact is available while the chat stays under the context window. Open a new chat and it is gone.
• With summaries, it survives only if the summary kept it. A summary written to capture destinations and dates may drop the dietary note.
• With saved facts, it is there in every chat and you can correct it when plans change, provided the assistant saved it in the first place.
• With retrieval, it comes back when a question about restaurants happens to match the passage where you said it. A question about museums may never surface it.

The same fact can be lost in four different ways, which is why it helps to know which approach an assistant uses.

What goes wrong with memory

• Wrong memories, stated confidently. A saved note can be mistaken, outdated, or taken out of context, and the assistant may still rely on it.
• Missing memories. Retrieval can fail to surface something that was stored.
• Contradictions over time. Preferences change, and an old note and a new one can disagree.
• Privacy. Stored memory is stored data. Who can see it, how long it is kept, and whether you can delete it all matter.

What to ask of an assistant with memory

• Can I see what it remembers?
• Can I edit or delete individual items, and delete everything?
• Is memory scoped to a project or topic, or is it one pool for everything?
• Does it tell me when it is using a remembered fact?
• What is stored, for how long, and who can access it?

If the answers are vague, treat memory as a convenience, not as a record.

Where NAMELESS stands

Lasting memory is something NAMELESS is built for, alongside unlimited context. NAMELESS is still in training and not yet available. Our privacy policy is still a draft and will be published before NAMELESS leaves private beta; what memory stores and how you control it belong in that policy, and it will say so plainly.

You can join the waitlist to follow progress.