Context and Memory: Long Conversations, Documents and Custom Instructions

"I already told you that!" is a very human reaction, and many people have it with an AI assistant too. To avoid that frustration, you need to understand how a model "remembers" a conversation, what happens when it gets long, how to give it documents, and how to set preferences once and for all. This lesson explains how context really works.

So far you have written good prompts and refined them over several exchanges. Behind this smooth conversation lies a mechanism with real limits, and knowing it will make you much more effective.

The model has no memory, only a context

A language model does not remember your previous chats the way a colleague would. Every time you send a message, the tool assembles a block of text that typically contains:

• hidden instructions from the provider (the "system prompt");
• your custom instructions, if you set any;
• the whole conversation so far: your messages and its previous answers;
• your new message, and any attached files or retrieved text.

The model reads that entire block and writes the next reply. What feels like "memory" within a conversation is actually the tool re-sending the conversation each time. This block is the context.

The context window

The context has a maximum size, measured in tokens, called the context window. It must hold the input and the output together. Modern models can handle quite long contexts, but the window is never infinite, and its size varies from one model and plan to another.

When a conversation exceeds the window, something has to give. Depending on the tool, older messages may be dropped, shortened or summarized automatically. The model then simply no longer sees them, and it may contradict itself or forget a constraint you gave at the start.

Longer is not always better

Even within the limit, a very long context can cause trouble:

• Dilution. Important instructions get buried among many other messages, and the model may give them less weight.
• Contradictions. You change your mind along the way, and the earlier versions remain in the context, so the model may mix them up.
• Drift. After many exchanges, the style or focus slowly shifts away from what you originally asked.
• Slower, costlier use. More text to process, which can matter in tools with usage limits.

Providing documents

You can give the model material to work from: pasted text, an attached file, or, in some tools, a link or a connected folder. This is one of the best ways to improve accuracy, because the model can rely on your source instead of its general memory.

Good practices:

1. Say what the document is and what you want from it. "Below is the minutes of last week's meeting. Extract the decisions."
2. Mark where it starts and ends. Use clear separators so the model does not confuse your instructions with the document's own text.
3. Ask it to stay inside the document. "Use only the information in the text. If something is not mentioned, say 'not stated.'"
4. Ask for evidence. "For each point, quote the sentence it comes from." You can then check quickly.
5. Handle long documents in parts. If a document is longer than the context window, or if the tool seems to skip sections, split it and process section by section, then combine the results.

Custom instructions and system-level settings

Many tools let you set custom instructions (sometimes called a profile, preferences, or a system prompt) that are added automatically to every conversation. It is a way to say, once and for all, things like:

• who you are and what you usually do ("I am a project coordinator in a small association");
• how you want answers ("Plain language, short paragraphs, no unnecessary flattery");
• rules to always follow ("If you are not sure, say so. Never invent sources.");
• your language and formatting preferences.

Here is an example of a compact, useful set of instructions:

I work as an office coordinator in a small non-profit.
Write in clear, friendly English with short paragraphs.
When I ask for a draft, give one version, not several.
If you lack information, ask me before guessing.
Never invent figures, names or sources. Say "I'm not sure" instead.

Some assistants also have a memory feature that stores facts about you between conversations, and others let you create separate projects or workspaces, each with its own instructions and documents. These features save time, but they also mean that more of your information is stored. Review what is saved, delete what you do not want kept, and check the settings before assuming either way.

What belongs in custom instructions, and what does not

• Yes: stable preferences, tone, role, general rules for reliability.
• No: sensitive personal data, secrets, or very specific task details that change each time (put those in the prompt).

When to start a fresh conversation

A new conversation is not a failure; it is often the best fix. Start fresh when:

• the topic has changed completely;
• the model keeps repeating an earlier mistake or ignoring a rule;
• you have changed your mind several times and the history is a mess;
• the conversation is very long and answers are getting worse or inconsistent;
• you want an independent second opinion, uninfluenced by the earlier exchange.

Before you leave, ask for a handover summary:

Summarize this conversation for a fresh start: the goal, the decisions
we made, the constraints to respect, and what is still to do.
Keep it under 150 words, as a short list.

Check the summary, correct it, and paste it as the opening of a new conversation. You keep the useful decisions and drop the noise.

A note on privacy and shared tools

Because context is re-sent and sometimes stored, anything you put in a conversation (including documents) may remain in your history. If several people use the same account or workspace, they may see it. Think about what you are putting in, and remember that deleting a conversation does not always mean immediate erasure on the provider's side; the retention rules are in the terms of the service.

Your turn: manage context like a pro

Exercise 1. Write a set of custom instructions (five lines maximum) for your personal work assistant: who you are, how you want answers, and two reliability rules.

Exercise 2 (project). Take a document you are allowed to share (a public article or your own notes, with no personal data about others). Write a prompt that gives the document, restricts the model to it, asks for three key points with a supporting quote for each, and asks it to say "not stated" when something is missing.

I am a freelance project manager working with small businesses.
Answer in plain, friendly English, with short paragraphs.
Give one draft by default, and ask me questions if information is missing.
Never invent facts, figures or sources; say when you are unsure.
For long answers, start with a two-line summary.

Below is an article. Use only its content.
Task: give me the three key points.
Format: for each point, one sentence followed by a supporting quote
in quotation marks. If the article does not cover something I ask,
write "not stated" instead of guessing.

--- ARTICLE ---
[paste the text here]
--- END ---