The words behind long AI chats, defined
Short answer: most long-chat problems come down to two limits: the context window, which is how much the model can read at once, and the browser page, which is how much it can draw. The terms below explain both, plus the limits and tools around them.
The same few words come up whenever a conversation gets long: context, tokens, memory, limits. They are easy to mix up, and mixing them up leads to the wrong fix, such as buying a bigger plan for a problem in the browser tab. Each definition here is short enough to quote.
- Context window
- The amount of text a model can read at once: the conversation so far, attached files and its instructions. It is measured in tokens. When a chat outgrows it, the oldest parts fall out of view.
- Token
- The unit a model reads text in. In English a token is roughly three-quarters of a word; code and many other languages use more tokens per word.
- Context limit
- The point where a conversation no longer fits the context window, also called the maximum conversation length. Some apps then refuse to continue and ask for a new chat; others quietly drop or compress the oldest turns.
- Context management
- Also called context handling: deciding what the model sees. One topic per chat, standing material in project or custom instructions, key decisions restated, and a handover rather than a transcript when you start over.
- Compaction
- Replacing older turns with a shorter summary so a conversation still fits the window. It keeps the gist and loses detail, often the exact constraint you needed.
- Memory
- A small set of facts an assistant saves about you and brings into new chats. It is separate from the context window: memory does not keep a long conversation in view.
- Custom instructions and projects
- Standing material an app adds to every chat in an account or project, so you do not paste it again. It still takes room in the context window.
- Handover
- A short message that starts a new chat with the goal, the decisions, the current state and the next question from an old one. It carries conclusions, not the transcript.
- Temporary chat
- A mode in which the platform does not keep the conversation in your history. It is designed not to persist, so it cannot be recovered afterwards.
- Usage limit
- The allowance a plan gives you before the app refuses or slows requests, also called a rate limit. It is often weighted by length and features rather than a simple count of messages.
- Rolling window
- A limit counted over the last few hours or days instead of per calendar day, so it lifts a set time after you were cut off, not at midnight.
- Lazy loading
- How chat sites keep a long conversation responsive: only part of it is kept in the page and the rest loads as you scroll. It is why Ctrl+F misses older messages.
- Speed engine
- Tvara's way of keeping a long chat fast: messages off screen are put to sleep, not deleted, and wake as you scroll to them.
- Minimap
- A thin map of the whole conversation on the edge of the page. Hover a mark to preview that message; click to jump to it.
- Local archive
- A copy of your conversations kept on your own computer. It can be searched offline and survives when the provider deletes a chat.
- Export and backup
- An export is a file for reading, such as Markdown, HTML or JSON. A backup is a file for restoring: complete, and ideally encrypted. You want both.
Common questions
What is the difference between memory and context in ChatGPT?
Context is the conversation the model can read right now, limited by its context window. Memory is a small set of saved facts carried into new chats. Memory does not make a long conversation fit.
Why do long chats use up usage limits faster?
Every message sends the conversation back to the model. On plans metered by usage rather than message count, such as Claude's, a message late in a long chat costs more than the same message in a new one.