Memory

The model keeps a memory of useful context from your conversations, so you don't have to provide the same information again. You control what goes into that memory through a simple feedback loop: give a thumbs up or thumbs down on any response the model generates.

For example, if you ask the model to summarize an article from an NVIDIA blog post, you can tell it to remember the summary within the project scope by clicking the thumbs up icon on that response.

This indexes the information into memory. The next time you ask a question about Diffusion Gemma, the model can answer without fetching the blog post again.

Curating memory this way makes results more predictable — responses reflect your style and the information you've chosen to keep, which leads to better generation of new content.

Giving a thumbs down does the opposite: it removes the information from the model's context.