🗂️ FileGram — AI That Learns Your Habits from How You Handle Files

What if your AI assistant could learn your preferences not from what you say, but from what you actually do with your files?

We all have patterns — how we name files, organize folders, multitask across documents. These behavioral traces reveal more about us than any conversation.

Researchers led by Shuai Liu introduce **FileGram**, a framework that personalizes AI agents by observing file-system behavioral traces rather than relying on dialogue history.

The system has three core components:

• **FileGramEngine** — generates realistic, large-scale multimodal training data by simulating diverse file workflows, solving the privacy problem of collecting real user data

• **FileGramBench** — a diagnostic benchmark testing AI memory across four dimensions: profile reconstruction, trace disentanglement, persona drift detection, and multimodal grounding

• **FileGramOS** — a bottom-up memory architecture that builds user profiles from atomic file actions through three channels: procedural, semantic, and episodic memory

Imagine an assistant that knows you always name files "date-project-version" and automatically organizes new saves into the right folder with the right format — without you saying a word.

Experiments show that even state-of-the-art memory systems struggle with FileGramBench, proving this is still an open challenge. But FileGramOS significantly outperforms dialogue-based approaches.

The project is fully open-sourced on GitHub.