Visual explanationThere is a boundary for what content can be referenced this time
Instructions and historyTask requirements, application instructions, and selected conversations
Data and tool resultsRelevant chunks actually read and included
Reserved for generationSubsequent output also needs available space
Context budget allocation (teaching illustration)

Saving or uploading materials does not mean the entire text enters this context.

Composition is for illustration, not drawn to scale with actual usage. Applications can first filter, retrieve, or summarize materials; different models have different output and inference usage rules.

Context: Information the model can reference when processing the current task, such as requirements, conversation, document excerpts, and tool outputs.

Why the Same Request Can Yield Different Results

"Prepare a Silk Road lesson," paired with "Grade 7, 40 minutes, haven't systematically studied the geography along the route," will be closer to the actual goal than the title alone. Adding M01–M03 documents, the model can also organize content around these references.

These requirements, background, and documents can become part of the current context. Context may also include application instructions, previous conversation, and tool outputs—not just the user's final message.

Saving, Uploading, and Reading Are Three Separate Things

StateCan indicateCannot yet prove
File is on your computerThe file exists on the deviceThe application can access it or the model has read it
File has been uploadedThe application received the fileThe entire file content entered the current input
Tool reads back the fileA specific read operation succeededThe application provided the complete file verbatim to the model

When using this, you can have the system point to the passages or document numbers it cited, then verify against the original text. A claim of "already read" alone is insufficient to confirm that all documents were actually used.

How Much Material to Provide

Select information based on the task: student background controls difficulty, M01 and M02 are used for instruction, M03 for discussion. Irrelevant notifications need not be included; example formats and factual evidence should each be identified by purpose to avoid copying old content wholesale.

When material is lengthy, mark relevant sections, core questions, and conditions that must be preserved. Do not assume that the longer the conversation goes on, the more completely every earlier statement remains available.

Going Deeper: Capacity and Filtering

The model has a capacity limit for information it can process at once, typically measured in tokens. Applications may truncate, retrieve, compress, or summarize material—the specific approach varies. Model output and some reasoning steps also consume the corresponding budget, so you cannot gauge capacity solely by uploaded file size.

Memory refers to information an application may persist across conversations; it only enters the current context when retrieved and used for the present task. Long-term preservation and current visibility are not the same concept.

Next Steps

Learn about Memory and Tokens, or read How to Provide Materials.

Reflect

Does having a file on your computer, uploaded, and read all mean the model can reference all the material this time?

Reference judgment: No; you need to confirm which information was read and selected—uploading and saving do not equal complete entry into the current context.

Source and Scope

Reference OpenAI · Conversation state. Technical specifications verified on 2026-09-09; examples and instructional organization written by this site. Specific product support scope and operation methods should be verified separately; this article makes no cross-product feature commitments.