ChatGPT Context Window: What Fits and What Gets Lost
The context window is the working information available for a response. Learn how it differs from saved memory and why a long chat needs a clear current brief.
Updated 6 min read
ChatGPT’s context window is the working information a model can use for a response, measured in tokens, small units of information. It has to accommodate the task’s input and the generated answer; it is different from a saved conversation or memories available across chats.
That distinction matters when a project includes long files, many revisions, and decisions scattered through a chat. A tool like fenbs keeps the current work record outside ChatGPT, so a new session can read the goal, decisions, and next step without reconstructing every earlier exchange.
What does the context window actually hold?
Think of the context window as a working desk. The important question is which instructions and evidence are available for this task. A document stored elsewhere is more like a reference on a shelf: its existence does not prove the relevant page is on the desk.
OpenAI’s context window documentation (opens in a new tab) describes the budget for input, output, and reasoning tokens used by some models. Those are model concepts; the ChatGPT application decides how it prepares a particular conversation and its sources for the model.
| Information | What it is for | What it does not prove |
|---|---|---|
| Current request | States the outcome you want now. | That all earlier instructions agree with it. |
| Conversation material | Helps interpret follow-up questions and prior choices. | That an old idea remains approved. |
| Source material | Provides facts and evidence for the task. | That every uploaded page informed this answer. |
| Saved memory | Can carry useful context across conversations. | That an exact source or full transcript is present. |
For the general concept across assistants, read the guide to an AI context window. Here the practical concern is how to keep a ChatGPT project understandable as the conversation and source collection grow.
Is saved memory extra space for the current chat?
Treat memory and context capacity as different ideas. Memory concerns carrying useful information from earlier work into future work. Context concerns what the model can use for this response. Saving a detail does not turn a finite working capacity into an unlimited one.
OpenAI’s Memories documentation (opens in a new tab) describes memory as a way to carry useful context forward and cautions against making it the only home for required guidance. If a decision must control today’s output, make its current wording and authority explicit.
For example, “the team prefers short proposals” is different from “this proposal must include the approved assumptions table.” The first is background. The second is a requirement you should be able to point to and verify in the finished result. In fenbs, people’s accepted Decisions and rules remain with the project for connected assistants to read.
Our guide to ChatGPT project memory covers that narrower feature. Do not use remembered context as evidence that ChatGPT has read the particular document or version you need for the current task.
How do files and long conversations affect what fits?
A request can involve more information than its latest sentence. “Compare these revisions” needs the revisions; “continue from the approved outline” needs the correct outline and its status. Relevant prior messages and source content can therefore make a seemingly short follow-up a substantial task.
OpenAI’s Projects guidance (opens in a new tab) explains how project chats can share uploaded files, instructions, and connected sources. That organization helps you return to related work. It does not mean that adding a file makes its entire contents a permanent part of every answer.
Separate accessibility from actual use. Ask which source and section support an important answer. If a question requires reading every entry in a large document, state that requirement and request a result you can check against the document, such as a list with page references.
The output also needs room. Asking for a complete long report after supplying extensive material is a different task from requesting a short summary. Plan the analysis and deliverable together rather than assuming the input capacity is an equal promise about answer length.
Does a missed detail mean the window is full?
No. A wrong answer alone does not diagnose a capacity problem. The source might be missing, ambiguous, obsolete, or simply misread. The instructions might also conflict. Check these ordinary causes before deciding that a larger window would solve the issue.
- Ask which source and passage support the disputed answer.
- Check whether that source is the current version and actually available.
- Restate the task and the specific constraint that was missed.
- Try the question with a focused excerpt that includes its necessary definitions and exceptions.
If the focused version succeeds, you have a useful direction for improving the working context, not proof of a particular internal failure. Avoid asking the model to certify that it remembers everything. Demonstrating the correct decision and supporting passage is a stronger practical check.
Use ChatGPT context limits for capacity questions and context window management for the broader maintenance workflow. A published model number is only one part of deciding how to prepare this task.
What belongs in a fresh project brief?
A restart brief should explain the current state without making the next session relive the whole discussion. Include the goal, accepted decisions, relevant source versions, unfinished work, and next action. Keep a qualification beside the fact it qualifies.
Say a design studio has reviewed three versions of a proposal. The useful handoff might say that the second structure is approved, the introduction has been revised, and the delivery dates remain unverified. “The proposal is approved” would be shorter but would lose the distinction that matters next.
Have the assistant draft a restart summary, then compare it with the source decisions. Preserve uncertainty and links to the original evidence. A summary should direct the next session to the right work; it should not replace an exact source when the task requires one.
That is the role of the fenbs handoff workflow: save the current state so the next connected assistant can start from a reviewed record. It helps whether the conversation changed because it became long or because a different person took over.
Keep the current brief outside the long conversation
Add the active project to fenbs and record the decisions that govern its next task. Ask your connected assistant to read Rules and Where we left off at the start, then save changed progress before stopping. Begin with one task whose current state you can check today.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs?
- Create a free account, then add your project and current tasks to a kanban board. It is free for one person, with no credit card.
- Record accepted choices in Decisions and rules, keeping open questions distinct from decisions a person has made.
- Add the context notes and source links needed for the next step. Keep long original documents in their appropriate home.
- Follow the integration instructions to connect ChatGPT over MCP, a standard connection for using an outside tool. Ask it to read the project context before work and save progress before it stops.
fenbs does not enlarge ChatGPT’s context window, inspect its internal context, or import its chats. It fits ongoing work that needs a durable project record. For a self-contained question with one small source, a focused conversation may be enough.
How should you compare capacity across assistants?
Compare the actual product and model you intend to use, then test a representative task. Anthropic’s Claude context guidance (opens in a new tab) is the place to check its changing capacities. A larger number does not by itself establish better source selection or a more accurate answer.
Keep your reviewed brief and evidence usable whichever assistant you choose. The context engineering hub collects the next steps: select relevant material, make current instructions clear, and verify the result against its sources.