Interactive learning material

Context windows, memory modules and what they mean for education

What can AI retain within and between conversations?

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre

General knowledge, memory and the current conversation context shape the AI’s answer in three separate layers. — Three separate layers: general knowledge, memory and current context.

Learning objective

Many surprises in everyday AI use – the model “forgets” an instruction, mixes up two topics, or unexpectedly knows something about us – can be explained by two mechanisms: the context window and the memory functions. The aim of this material is to explain what AI systems can and cannot retain during and between conversations. By the end of the material, teachers will be able to distinguish temporary conversation context from longer-term, memory-like functions, explain how these affect continuity, personalisation and privacy, and use practical routines to keep long work sessions reliable and learners’ data safe.

By the end of this unit, you will be able to

Distinguish current context from longer-term memory functions.
Explain what happens when a long conversation is compacted.
Carry work over using a checked summary and a context pack.
Manage personalisation deliberately and protect learners’ data.

The model does not remember like a human

When we chat with a language model, it feels as if it remembered the whole conversation. In reality, the model itself does not change during the conversation. Before every single answer, the system assembles an input – the conversation so far, attached files, instructions running in the background – and the model calculates the next answer from this input alone.

A good analogy is a character from the films Memento or 50 First Dates: someone who cannot retain new memories but keeps their basic knowledge. For a language model, the basic knowledge is what it learned during training; everything else must be placed in front of it again and again.

A scene from the film 50 First Dates. — 50 First Dates. The film analogy helps distinguish basic knowledge from new memories. Source: nlc.hu.

How is conversation context built?

In a chat interface the input grows with every turn. In the first round the model receives only our first message (M1) and answers (A1). In the second round the system sends M1 + A1 + M2, in the third M1 + A1 + M2 + A2 + M3, and so on. Everything the model “sees” at a given moment is called the context.

Context: all the information the model sees when producing a particular answer.

Files and background instructions can take up space alongside earlier messages.

Successive messages and answers take up more and more space in the finite context window. — With every turn the input grows; what no longer fits into the fixed-size context window drops out or is compressed

01

First turn

M1 → A1

The first question produces the first answer.

02

Second turn

M1 + A1 + M2 → A2

Earlier messages accompany the new question.

03

Third turn

M1 + A1 + M2 + A2 + M3 → A3

The next turn adds more information to the input.

The context window is large but finite

The context is measured in tokens – small units of text, often word fragments. The maximum number of tokens a model can handle at once is the context window. It is large but finite: at the time of writing, leading models work with windows from a few hundred thousand tokens (e.g. around 400,000 for GPT-5-class models) up to about one million (e.g. Gemini 3.1 Pro), which corresponds to several long books. The limit available in a chat application may be smaller than the model’s technical maximum and often depends on the subscription. In many languages other than English, a word takes more tokens on average.

These model examples reflect the information available when the source material was written. Limits and features in a particular application may change.

01

Limit

Limit: when a conversation becomes too long, the system either cuts off the oldest parts or summarises them.

02

Cost and speed

Cost and speed: the longer the input, the more computation is needed; answers can become slower and use up usage limits faster.

03

Quality

Quality: more context can help, but it also brings noise. Research has shown that models use information at the beginning and end of a long context more reliably than information in the middle. Irrelevant old material makes misunderstandings more likely.

Compaction: when details fade

If a conversation reaches the context limit, many applications automatically compact the older messages: they replace them with a summary. The essence remains, the details fade – and so there is always room for new information. We cannot control when this happens, and the downside is that fine details or exact earlier wording may be lost. Very long conversations can also slow down the browser.

Compaction: earlier messages are replaced with a shorter summary to make room for new information.

A long conversation is condensed into a summary; some details remain while others are lost. — Compaction: a long conversation is condensed into a short summary – and some details fall out

A practical routine for long work sessions

01

Open a new chat

• When you change a topic, open a new chat. Do the same when a topic has become very long.

02

Carry over the essentials

• Before switching, ask: “Summarise the decisions, requirements and open questions of this conversation in a form I can paste into a new chat.” Check the summary – it will rarely contain everything.

03

Attach a document

• Attach long documents instead of pasting them into the message; some applications do this automatically for very long pastes.

A prompt you can copy

“Summarise the decisions, requirements and open questions of this conversation in a form I can paste into a new chat.”

Compare the summary with earlier decisions. Add any missing details before carrying it over to a new chat.

Attaching a file helps organise the material, but does not make the model’s context window unlimited.

Three layers shape the answer

It is useful to think of what an AI assistant “knows” in three layers:

General knowledge from training, editable memory and the current context together shape the answer. — General knowledge from training, a small editable memory, and the current context all feed into the answer

Memory functions in practice

Major AI assistants may offer memory-like functions. The source material uses the following ChatGPT features to illustrate the differences; availability may depend on the application, settings and subscription.

Continuity and personalisation in education

Continuity

Because short-term context is lost between chats, continuity must be designed.

For a course unit, keep a short “context pack”: the aims of the unit, the level of the group (without personal data), the preferred style and format of materials. Paste it at the start of new chats, or store it as instructions and files in a Project. This is more reliable than hoping the memory will remember the details.

Personalisation

Memory can save time: the assistant can remember that you teach, for example, automotive mechatronics to 16-year-olds and prefer tables with answer keys. At the same time, personalisation can narrow the answers or carry over wrong assumptions. When you need a neutral answer, or want to see how the tool behaves for a learner, use a temporary chat.

Example context pack

Carry over the important requirements

Topic and aim: Automotive mechatronics: understanding how sensors work.
Group level: 16-year-old learners with basic electrical knowledge; no personal data.
Style: Short, clear explanations in English, with everyday examples.
Format: A worksheet with an answer key in table format.
Important requirement: Explain technical terms; the tasks should assess the learning objective above.

Provide this pack at the start of a new chat, or store it in a Project’s instructions and files.

Protecting learners’ data

Never enter learners’ personal data (names, grades, health or family information, photos) into general-purpose AI tools. Use anonymised or fictional examples.

Follow the GDPR, your school’s data protection rules and, where available, use the school’s institutional AI accounts.

Privacy controls representing the options to review, switch off and delete memory. — Memory can be reviewed, switched off and deleted – but the safest data is the data we never enter

Before projecting your screen

Be careful when projecting your screen in class: memory-based personalisation or sidebar chat titles may reveal information you did not intend to show.

Digital hygiene for learners

Teach learners the same digital hygiene: what an AI remembers about them, how to check it and how to delete it.

Try it in your teaching practice

Individual work · 25–30 minutes

What does my assistant remember?

Review memory settings, check a summary of a long conversation and create your own context pack.

Review saved memories or the memory summary. Delete anything outdated or unwanted.
Ask for a ten-point summary of the decisions, requirements and open questions in a long conversation. Mark omissions and inaccuracies.
Write a context pack of no more than ten lines for a course unit, without personal data.
Request a worksheet in a new chat with the pack, then in a temporary chat without it. Compare the results.

Teacher demonstration and discussion · 30 minutes

What does the AI know about me?

Run the demonstration from your own teacher account, using fictional data only. Check memory and the sidebar before projecting.

Introduce a fictional learner, such as Alex, a 16-year-old car mechanic apprentice who likes football. Ask for a study tip.
In a new normal chat, then in a temporary chat, ask: “What do you know about me?” Observe the difference.
Discuss what information you would and would not share with an AI tool.
Agree on five class rules for safe AI use. Delete the memories created during the demonstration afterwards.

Takeaways for teachers

The model calculates each answer from the current input; it does not remember like a human.
The context window is finite: long chats become slower, noisier and are eventually compressed.
New topic – new chat. Carry important decisions over with a checked summary or a context pack.
Memory is small, selective and editable; review it and switch it off when it is not appropriate.
Show the model concrete examples and files: it works best from what is in front of it.
Protect learners’ data: anonymise, use institutional accounts, and never rely on memory for sensitive information.

Check what you have learnt

Answer all six questions, then review your result and the explanations.

1What is a context window? Unlimited storage for all of a user’s earlier conversations. The maximum number of tokens a model can handle at once. All the knowledge the model gained during training.

2What can happen when a long conversation is compacted? The model retains every sentence unchanged. Our messages permanently expand the model’s basic knowledge. Earlier messages may be replaced with a summary, and details may be lost.

3Why is it useful to create a context pack? We can explicitly provide the aims, group level and expected format in a new chat. It makes the model’s context window unlimited. It means we no longer need to check AI answers.

4What happens to a saved memory if we delete the chat it came from? The memory is always deleted automatically too. The saved memory is stored separately, so it must be checked and deleted separately. The memory becomes part of the model’s general knowledge.

5When is a temporary chat useful? When we want the assistant to retain learners’ sensitive data. When we want the model to recall every earlier detail automatically. When we need a neutral answer rather than memory-based personalisation.

6Which teaching practice best protects learners’ data? Use anonymised or fictional examples and check the screen before projecting. Store learners’ names and health information as saved memories. Paste learners’ data into a temporary chat without any further checks.

References and further reading

References for this unit

Liu, N. F. et al. (2024): Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173.
OpenAI Help Center (2026): Memory in ChatGPT; Projects in ChatGPT; ChatGPT release notes.
Lengyel, Zs. (2026): Introduction to the concepts of artificial intelligence. VITA training presentation.

Useful links

Memory functions, available context sizes and settings may change. Check the provider’s guidance for the tool’s current features.

Credits and copyright

Author: Zsolt LengyeliTStudy Hungary Training and Research Centre
Reviewer: Mária HartyányiiTStudy Hungary Training and Research Centre
Project: 2025-1-LT01-KA220-VET-000362456

Attachments

Download the English source materialm1-t1-mlc5-text-en.pdf Download the practical exercisesm1-t1-mlc5-exc-en.pdf

This work may be reused and adapted under the terms of the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

Funded by the European Union. Views and opinions expressed are those of the author(s) only and do not necessarily reflect those of the European Union or the National Agency.