The Useful Question
English

Why does AI miss something you already told it?

An earlier detail can be absent from the material used for the next answer, or present but poorly used. A long conversation and a memory feature do not guarantee that every requirement will carry through. The practical repair is to restore the relevant evidence, agree on a short current brief, and check the result against it. Start by separating what the chat displays, what the model receives, and what the answer actually follows.

First, make the missing requirement visible

Imagine preparing a flyer for a neighborhood repair café. Early in the chat, you specify black-and-white printing and no photographs. After several rounds about the headline and layout, the AI proposes a full-color photo banner. You can still scroll up and see your original instruction.

This is a fictional teaching scenario. The proposed answers and revisions are simulated; no live model experiment was performed. It illustrates why “I told you already” is a reasonable complaint but an incomplete diagnosis.

Try a focused repair: “The flyer must print in black and white and contain no photographs. The venue has changed to the community hall. The date is not confirmed. List those current requirements before revising, and do not invent a date.”

Then inspect the revised flyer. Restating the rules correctly and following them in the finished output are separate checks. A polished requirements list is useful only if it helps you catch what the draft still gets wrong.

The visible conversation is not the context window

A model generates an answer from the input available for that request. Its context window limits how much tokenized material the request can use. Tokens are processing units, not a fixed number of English words or Japanese characters. OpenAI's conversation-state documentation also explains that input and generated output, including reasoning tokens where applicable, share the context budget.

A chat application manages how conversation material reaches the model. As a conversation grows, it may select or compact earlier material; implementations differ. OpenAI documents compaction as an API mechanism for carrying prior state forward with fewer tokens. That is an example of context management, not proof of what happened inside your particular chat.

The useful distinction is between a transcript retained by the application and the material used in a particular response. Being able to scroll to a sentence does not establish that the same sentence was supplied verbatim at that moment. Without application-level evidence, do not diagnose every omission as “the window filled up.”

A visible conversation and optional remembered information can contribute to the current input. Two separate checks follow: whether the required detail is available, and whether the answer follows it.
Original editorial illustration
Expand the diagram

Scroll within the diagram horizontally; arrow keys work when focused.

A visible conversation and optional remembered information can contribute to the current input. Two separate checks follow: whether the required detail is available, and whether the answer follows it.
Open the diagram at full size

Figure: An original conceptual map of two failure points. It does not disclose or assert the implementation of any particular chat product.

A detail can fit and still be missed

Capacity is one issue; using the material effectively is another. In the historical Lost in the Middle study, researchers changed where relevant information appeared in long inputs. Performance varied with its position in the models and tasks they tested, including difficulties with material in the middle.

That study was published in 2024. It does not establish a failure rate for a 2026 product, prove that every model has the same pattern, or reveal the cause of your flyer mistake. It provides a useful warning against treating a large advertised context window as perfect recall.

A conflicting instruction can also need resolution. Perhaps an early draft used the library as the venue, while a later message moved it to the community hall. A recap that blends both is not a reliable working brief. Marking the hall as current and the library as superseded makes the intended change explicit, regardless of why the model mishandled it.

Memory has a different job

A memory feature can carry selected personal context across conversations. It is not the same thing as keeping every earlier sentence available for every answer. OpenAI's current Memory in ChatGPT guide says memory does not retain every detail; available features and controls vary by plan, region, platform, and workspace settings.

A preference such as “I like concise drafts” may be useful across projects. “This flyer must use the newly approved venue name” belongs in the current project brief as well. For wording that must be reproduced exactly, keep the approved source and provide the relevant passage when needed.

Ask the product to review remembered information when the feature supports it, and correct outdated details. Still keep a separate authoritative copy of important project decisions. Personalization can reduce repetition; it is not a substitute for a document you can inspect and approve.

Build a short handoff you can check

When a conversation has accumulated several revisions, create a compact brief containing:

For the flyer, that means the community hall, black-and-white printing, no photos, and an unconfirmed date. It does not mean repeating every rejected headline. Ask the AI to prepare the brief, but compare it with the original decisions yourself before treating it as approved.

If you start a fresh conversation, bring that checked brief and the relevant source material. Starting over without them can discard useful context. Staying in the existing chat and restating the brief can also be reasonable. The important change is making the current requirements explicit and inspectable.

Check the output before adding more history

Compare the next draft against each requirement. Does the layout remain usable in black and white? Has a photograph returned? Is the venue correct? Is the date visibly unresolved? These are concrete acceptance checks, not a request for the AI to rate its own confidence.

If the same rule keeps being missed, narrow the task to the offending section and supply the exact source again. Adding a whole archive can make review harder without clarifying the requirement. The aim is a smaller, reliable working brief and a verifiable result, rather than an endlessly growing conversation.

Sources and limits

Part of “How an AI answer is made”; readable on its own. Primary sources were checked on October 11, 2026. This is not a hands-on recall benchmark. No universal turn limit, model-specific capacity, or guaranteed recovery method is claimed; the cause of an individual omission remains unverified without evidence from that application.