Researchers built a synthetic dataset to train AI assistants that remember how you actually work

A team created FileGram, a system that generates fake file-system activity patterns to train AI agents without collecting real user data. This solves a hard problem: teaching AI assistants to understand your work habits without needing access to your actual files or asking you to label your behavior.

Right now, personalized AI assistants that live inside your computer can't actually learn from your real work because privacy rules and the cost of collecting actual data make it infeasible. FileGram sidesteps that by generating synthetic traces that mimic realistic work patterns, then training memory systems on those fakes instead. This means the next generation of file-system agents could understand context and habit without the surveillance problem — or it could just mean the researchers built a clever benchmark that doesn't actually work when the AI meets real users.