Gemini memory?

[SebiXibes]

Hi, thanks for this amazing implementation!
I'm new to this, but am having a lot of fun experimenting.
One question I have, what happens to gemini's memory when it is being called? is each new invocation creating a new context window? or does it just go into its massive context window without "purging"? Apologies if this was written somewhere, but couldnt find it...

[jamubc]

The ask-gemini tool wraps Gemini CLI with the -p flag, this is a non-persistent session thus it disappears when the process finishes.
So yes, exactly how you've stated that each invocation is a "new" context window. It's a new "session" of Gemini.

There a couple workarounds that I am aware of, Gemini CLI can store memories via #something to remember messages, these persist across invocations and are typically stored in the working directory context such as a .gemini or GEMINI.md.

Since these memory files are typically stored in a project or in the ~/.gemini dir, new invocations should automatically read GEMINI.md. This in a sense gives you the ability to define rules or plans, styles, workflows, ect.

There is definitely more to explore in terms of continuation and resuming sessions, and it does sound possible.

In terms of what specifically happens on the Gemini side, you should look at gemini-cli repository or find changes in gemini related files when using Gemini.

[SebiXibes]

Thanks, i was a bit confused about this since it seemed that claude was trying follow-up prompts with the assumption that gemini would still remember the context. so it seems with a follow-up prompt you need to be explicit to claude that gemini doesnt have the full context anymore

[jamubc]

Yeah, If you want Claude to "continue" the previous conversations with gemini you need to work through the #memory syntax that Gemini-CLI has, just convince Claude to use it to let gemini "remember" things.

Claude will remember things, and can run an instance of Gemini for the sole purpose of "storing" a Gemini #memory

You might also have more luck forcing this behaviour with prompts such as: /ask-gemini /memory add always use snake case for variable names . Just tested in persistent mode if the tool prompt input is /memory add ..., it does persist.