Why does AI forget what you told it?

You told the chatbot your name, your plans, your dog's name, and a week later it has no idea. Here is how AI memory really works and how to make it remember what matters.

Dear my… editors · Editorial team

· 3 min de leitura

6 comentários
Nesta página

Short answer: A chatbot does not learn from your conversations the way a person does. Within one chat it can only see a limited amount of recent text, called its context window. When a chat ends, or gets very long, older details fall out of view. Some apps add a separate memory feature that saves a few notes about you, but it is selective and can be switched off. So forgetting is usually not a bug; it is how these systems are built.

The model itself does not change when you talk to it

The AI model behind a chatbot is trained once, on a huge amount of text, and then frozen. When you chat with it, it is not rewriting itself or filing your details away. Every time it answers, it simply reads the conversation so far and predicts a reply.

That means everything it "knows" about you in a chat comes from the text currently in front of it.

The context window: a desk, not a filing cabinet

Think of the context window as a desk. Everything on the desk can be used: your messages, its replies, any files you uploaded. Modern models have big desks, often enough for hundreds of pages. But the desk still has edges.

  • In a new chat, the desk is empty. Unless the app has a memory feature, the model starts from zero.
  • In a very long chat, old pages fall off. Many apps trim or summarize the oldest parts so the conversation can continue, and details get lost in that process.
  • Even on the desk, the middle gets overlooked. A well-known 2023 study, "Lost in the Middle", found that models use information at the start and end of a long text much better than information buried in the middle. Newer models are better at this, but the effect has not fully gone away.

Memory features: a few sticky notes

To feel more personal, apps such as ChatGPT, Claude and Gemini offer memory or personalization features. They typically work like sticky notes the app puts on the desk at the start of each new chat:

  • the app decides, or you tell it, what is worth saving (your name, your job, that you are vegetarian)
  • those notes are added to new conversations behind the scenes
  • you can usually view, edit or delete them in settings, or turn memory off entirely

These notes are short summaries, not a full record. That is why an assistant might remember that you have a dog but not what you said about the vet last Tuesday.

Why not just remember everything?

  • Cost and speed. Reading a longer history on every reply takes more computing power and time.
  • Accuracy. More text means more chances to mix up old and new information, for example a job you have since left.
  • Privacy. Many people do not want a permanent record of everything they have ever typed. Selective, editable memory is a compromise.

How to make AI remember what matters

  1. Keep related topics in one chat. Coming back to the same conversation keeps the history on the desk.
  2. Start long tasks with a short brief. "I am planning a 5-day trip to Lisbon in May with two kids, budget around €2,000" saves the model from guessing.
  3. Repeat key facts when they matter. If a detail from much earlier is important, mention it again.
  4. Ask for a summary, then reuse it. At the end of a long session, ask "Summarize what we decided" and paste that into the next chat.
  5. Use memory or custom instructions on purpose. Tell the app explicitly what to remember, and check the saved notes now and then.
  6. Do not store what you would not want saved. Passwords, ID numbers and sensitive health details are better left out of any chat.

A note on companions

AI companion apps, including Dear my, put extra effort into remembering the people and moments that matter to you, because a relationship without memory feels hollow. But the same rules apply underneath: memory is a set of saved notes, it can be imperfect, and you should be able to see and control what is kept.

Based on published research and product documentation as of October 2026.

Fontes

  1. Liu et al.: Lost in the Middle: How Language Models Use Long Contexts (2023)
  2. Anthropic docs: Context windows
  3. OpenAI Help Center: Memory FAQ

Escrito pela nossa equipe editorial com ajuda de ferramentas de IA. Viu um erro? Conte para nós na discussão.

Este artigo ajudou você?

Discussão6 comentários

Participe da discussão

Entre ou crie uma conta grátis para comentar. Só o seu apelido aparece.

EntrarCriar conta
  1. zoelxx

    lowkey want it to remember everything fr but also lowkey terrified of that 💀 like imagine it knowing every dumb thing i typed at 3am

  2. lbrandt

    My preference is clear: forgetting by default, with memory as an explicit opt-in. The article is right that a permanent record of everything one has typed is not something many people want. However, I recognise the convenience argument; a short, editable list of facts seems a reasonable compromise.

    1. zoelxx

      ngl @lbrandt that makes sense but opt in sounds like so much work 😭 i js want it to know my name and not my whole life

    2. lbrandt

      @zoelxx

      It need not be much work, however. According to the article, saved notes can usually be viewed, edited or deleted in settings. A brief review once a month would be sufficient, and one can also turn memory off entirely.

  3. Tomek

    I am using ChatGPT memory since few months and it is exactly like the article says lol. It remembered that I have dog but not the whole project I explained last week (sticky notes, jo?) The desk and sticky notes explanation on dearmy was better than what I find on random blogs 😄

    1. gracieo

      😂😂 the sticky notes thing is so true abeg!! mine knows my name but forgets the whole project sha!!! @Tomek now I ask for a summary at the end of the chat and paste it in the next one, it saves me!! 🙏🔥