Researchers have released FileGram, an open-source framework that enables AI agents to achieve personalization through file-system behavioral traces rather than dialogue interactions. Published on arXiv on April 6, 2026, the comprehensive framework addresses severe data constraints and privacy barriers that have limited personalization in AI coding and productivity agents.
File-System Operations Reveal Dense Behavioral Signals
Existing personalization methods are interaction-centric, relying on dialogue summaries that overlook dense behavioral traces in file-system operations. FileGram instead builds user profiles from atomic actions—file creates, edits, moves, and deletes—and content deltas showing what actually changed. This bottom-up approach captures work patterns and coding styles without requiring conversational data or violating privacy constraints.
Three-Component Framework Enables Research and Development
FileGram comprises three components. FileGramEngine is a scalable persona-driven data engine that simulates realistic workflows and generates fine-grained multimodal action sequences, overcoming privacy barriers through synthetic but realistic data. FileGramBench provides a diagnostic benchmark grounded in file-system behavioral traces, evaluating memory systems on profile reconstruction, trace disentanglement, persona drift detection, and multimodal grounding—remaining challenging for state-of-the-art systems. FileGramOS offers a bottom-up memory architecture that encodes traces into procedural, semantic, and episodic channels with query-time abstraction for efficient retrieval.
Bottom-Up Architecture Outperforms Dialogue Summarization
Experiments demonstrate that FileGramOS is effective at building user profiles directly from atomic actions and content deltas rather than dialogue summaries. By avoiding lossy LLM summarization, the architecture maintains fine-grained understanding of developer workflows. The framework was developed by Shuai Liu, Shulin Tian, Kairui Hu, Yuhao Dong, Zhe Yang, Bo Li, Jingkang Yang, Chen Change Loy, and Ziwei Liu, who have open-sourced it to support future research.
Implications for AI Coding Assistants
As AI coding agents like Claude Code, Cursor, and Windsurf become standard developer tools, personalization becomes critical for predicting needs, suggesting relevant files, and adapting to individual coding styles. FileGram provides infrastructure for building and evaluating such personalization without requiring dialogue data or compromising privacy.
Key Takeaways
- FileGram builds AI agent personalization from file-system traces—atomic actions and content deltas—rather than dialogue
- Framework includes data engine, diagnostic benchmark, and bottom-up memory architecture, all open-sourced
- FileGramBench evaluates memory systems on profile reconstruction, trace disentanglement, persona drift, and multimodal grounding
- Bottom-up approach avoids lossy LLM summarization while overcoming privacy barriers through synthetic data generation
- Addresses personalization needs for AI coding assistants like Claude Code, Cursor, and Windsurf