Training vs inference: what chats change

Intermediate · 5 min · Explain why a chat does not retrain a model, and what memory features store.

  • training
  • mental models

Training is how a model’s weights were set before you ever used it; inference is what happens each time you send a message. Your conversation runs inference — it does not retrain the model. Anything that feels like the assistant “learning” mid-chat is either the conversation still being in view, or a memory feature storing a note for later.

What your message actually does

Your words are turned into tokens, added to the conversation’s context window, and the model predicts a response from weights that were fixed at training time. That is the whole loop. Nothing is written back into the weights while you talk, so nothing you say can change how the model behaves for the next person — or for you, tomorrow, in a fresh chat.

What looks like learning (and what it is)

  • “It remembered my name.” Your name is still in the conversation. Ask in a brand-new chat and the effect disappears.
  • “I told it to always answer in bullet points and it does.” Custom instructions are text stored by the product and re-inserted into each new conversation. Useful — and editable evidence, not learning.
  • “It says it saved this to memory.” Some products keep a small store of notes and paste them into future chats. That is retrieval of stored text, not a changed model.
  • “My data made it smarter.” Providers may choose to collect conversations for future training rounds — but that is a policy decision written in their terms, not something your chat does automatically. It is a privacy question, not a mechanism, and it is worth reading the actual policy for the tool you use.

What happens to one message

  1. Your message is split into tokens and appended to the conversation's context.
  2. The model predicts a response from weights that were fixed before you arrived.
  3. The response returns; the conversation is stored as history by the product.
  4. A fresh chat starts with an empty context — which is why continuity has to be re-supplied or stored on purpose.

A bad example

Ask: “Remember this for next time: I always want sources with every fact.”

If the product has no memory feature, nothing is remembered. Next session starts blank, and the promise quietly expires the moment you close the tab.

A better example

Ask: “Store this preference where I can re-use it — or remind me to put it in my custom instructions.”

Now continuity has a home you can see and edit: a saved instruction, a reusable template, or a note you paste into the next chat.

Why it works

The distinction sounds academic until it changes behaviour. Knowing that context is the only memory in play tells you exactly what to do when continuity breaks: keep essential facts in the message, put standing rules in the product’s instructions field, and treat “I will remember that” as a sentence about the conversation, not about the model. It also sharpens privacy questions: whether your conversations are used for future training is decided by the provider’s policy, so read the policy — do not infer it from how the chat feels.

You chatted with an assistant for an hour yesterday about a project. Today, in a new chat, it asks what you are working on. Why?

Practice

Spot the training myths

A teammate drafted an explainer for your team wiki: “How the assistant learns from us.” You need it to be accurate before it starts shaping how colleagues think about privacy and data use.

Select every claim that misstates how training and inference work. Leave the claims that are accurate.

The AI answer

Every message you send is used to retrain the model overnight. The model's weights are fixed once training finishes, so a conversation cannot change how it behaves for anyone else. I have updated my knowledge with everything you told me. The assistant will remember your preferences in future chats because we enabled the memory feature, which stores short notes and re-inserts them into new conversations. When you open a new chat, it starts with an empty context, so anything you want it to know must be re-supplied or stored.
Select every part that should not be trusted as-is
Hint

Separate three layers: the weights (training), the conversation (context), and the product's notes store (memory).

Claims that attribute memory to the model itself — as if chatting wrote something in — are the ones to flag.

Transfer

  • Fresh chat, blank start: re-attach the brief, or use a saved instruction.
  • Standing rules (tone, format): put them where the product re-injects them automatically.
  • Essential facts: keep them in the message you send, not in a promise that they will be remembered.
  • Privacy questions: check the provider’s actual policy before trusting a chat summary of it.

Key takeaways

  • Training sets the weights; inference uses them. Conversations never change the weights.
  • Continuity comes from context, custom instructions or a memory store — all of it text, all of it editable.
  • Whether your data is used for future training is a policy question: read the policy, not the vibe.

Next

Finished this lesson?