Does Resuming a Saved Chat, Include Previous File Uploads?

[gregbythebeach]

When I resume a saved chat, does it also remember the files that I uploaded to it? Is it re-uploading them to itself or does it have it already saved even if they're really big files? How does that work?
I want to make sure that it's not just hallucinating when it continues the chat. Because when I resume it says it has 100% context, but I wouldn't think that's possible given that the files are about 500k characters each and there's 8 of them. When I first uploaded them for the original prompt, it usually uses a huge amount of the context.

They're on my desktop so I guess it was technically accessing them from my desktop originally I don't know if it's an actually upload or how that works.

[bdmorgan]

It is not loading all 4 million characters of your files into the context window each time. Your intuition is correct; that would exceed the context window of any current model. The "100% context" message refers to something slightly different.

Here’s a breakdown of how it works, which involves two key concepts: checkpointing and grounding.

1. Checkpointing: Saving the Conversation, Not the Files

When you have a conversation, gemini-cli saves the history of that conversation to a checkpoint file. This is what allows you to "resume" a chat. However, it doesn't save the entire content of the files you provided in that checkpoint.

Instead, it saves the conversation flow, which includes:

Your original prompt (e.g., "summarize these files: /path/to/file1, ...").
The model's response, which was likely a series of tool_call requests to read and process those files.
The results of those tool calls (e.g., summaries of the files, or confirmations that they were read).
Any subsequent questions and answers.
So, when you resume, gemini-cli reloads this entire transcript of events. The "100% context" message means it has successfully loaded this entire conversation history.

2. Grounding: Giving the Model Tools, Not Data

This is the core of how it handles large files. When you provide a file path in your prompt, you aren't actually "uploading" the file to the model in the traditional sense. Instead, you are grounding the model.

Here's what that means:

You give the model a file path: gemini-cli recognizes this is a local file.

The model gets a tool: The model is given access to file system tools (like read_file). It now knows it can read that file if it needs to.
Intelligent, On-Demand Access: The model doesn't immediately read the entire 500k characters. Instead, it makes an intelligent decision. For your initial request, it likely used its read_file tool to stream the content of the files, created summaries or analyses, and returned those to you.

History Contains the Result: The chat history now contains the result of that initial analysis. For example, it might contain a 2,000-character summary for each file. This is what gets loaded from the checkpoint.

How it Works When You Resume

When you resume the chat and ask a follow-up question, the following happens:

The full conversation history is loaded into the model's context. The model now "remembers" that it has already analyzed eight specific files and has its previous summaries available.
It still has access to the file system tools and the original file paths.
If your new question can be answered from the summaries already in the chat history, it will do so directly.
If you ask a very specific question that requires re-examining the source (e.g., "What is the exact value on line 2,345 of file X?"), the model can make a new tool_call to read_file again, but it might only read a small, specific chunk of the file this time.
In short: It's not hallucinating. It genuinely has access to the context of the previous conversation and retains the ability to re-read your local files at any time. It's a much more efficient system that uses tools to access data on-demand rather than trying to stuff massive file into its context window all at once.

Hope that helps...

[gregbythebeach]

Thanks for the detailed explanation. That's interesting and glad to know. It's just using great efficiency methods rather than hallucinating.

I was worried for a bit, as some other users indicated to me that they believe the AI would simply see the file paths in the resumed saved history of the chat, and thus the ai believes it still has memory of them, but really does not, and thus hallucinating with its responses.

Which i understand now is obviously not true

I tested what you said, with a new specific question "that requires re-examining the source (e.g., "What is the exact value on line 2,345 of file X?") and the context went down 20%, as expected.

I was also able to add new files (via a path), and it did not have too much of an impact on contact window or tokens.

[Manamama]

And if you dislike plain JSONs there, use: #3965