FileGram: Grounding Agent Personalization in File-System Behavioral Traces
| Authors | Shuai Liu et al. |
| Year | 2026 |
| HF Upvotes | 40 |
| arXiv | 2604.04901 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Coworking AI agents operating within local file systems are rapidly emerging as a paradigm in human-AI interaction; however, effective personalization remains limited by severe data constraints, as strict privacy barriers and the difficulty of jointly collecting multimodal real-world traces prevent scalable training and evaluation, and existing methods remain interaction-centric while overlooking dense behavioral traces in file-system operations; to address this gap, we propose FileGram, a comprehensive framework that grounds agent memory and personalization in file-system behavioral traces, comprising three core components: (1) FileGramEngine, a scalable persona-driven data engine that simulates realistic workflows and generates fine-grained multimodal action sequences at scale; (2) FileGramBench, a diagnostic benchmark grounded in file-system behavioral traces for evaluating memory systems on profile reconstruction, trace disentanglement, persona drift detection, and multimodal grounding; and (3) FileGramOS, a bottom-up memory architecture that builds user profiles directly from atomic actions and content deltas rather than dialogue summaries, encoding these traces into procedural, semantic, and episodic channels with query-time abstraction; extensive experiments show that FileGramBench remains challenging for state-of-the-art memory systems and that FileGramEngine and FileGramOS are effective, and by open-sourcing the framework, we hope to support future research on personalized memory-centric file-system agents.
Engineering Breakdown
Plain English
FileGram addresses a critical gap in personalizing AI agents that operate within local file systems by grounding agent behavior in file-system traces rather than traditional interaction logs. The paper proposes a framework with three core components: FileGramEngine (which simulates realistic workflows to generate multimodal action sequences at scale), FileGramBench (a benchmark for evaluation), and personalization methods that learn from dense behavioral traces in file operations. The key insight is that file-system operations contain rich, privacy-preserving signals about user behavior that can be used to train and personalize agents without collecting sensitive interaction data across organizations. This approach overcomes severe data constraints by enabling scalable synthetic data generation while maintaining strict privacy barriers.
Core Technical Contribution
The core novelty is reframing agent personalization from interaction-centric models to file-system-behavior-centric models, enabling privacy-preserving personalization without requiring multimodal trace collection across organizations. FileGramEngine introduces a persona-driven data simulation engine that generates fine-grained, realistic multimodal action sequences (file operations, timing, context) synthetically at scale, bypassing the privacy and logistical barriers of collecting real-world traces. The framework incorporates memory grounding mechanisms that anchor agent state and decision-making to actual file-system operations rather than abstract interaction histories, creating a more realistic and interpretable personalization signal. This represents a fundamental shift in how we approach agent personalization: instead of learning from what users say they do, learn from what they actually do in their file systems.
How It Works
FileGram operates through a three-stage pipeline. First, FileGramEngine synthesizes realistic user personas and simulates their workflow patterns (file creation, modification, access sequences, time intervals) to generate large-scale, labeled training data; this synthetic generation is parameterized by persona definitions that specify domain, role, and behavioral patterns. Second, the framework extracts dense behavioral features from file-system traces: operation sequences, file metadata (type, size, modification time), directory hierarchies, temporal patterns, and access correlations—creating a multimodal representation that captures both what users do and how they do it. Third, personalization is learned by training agent memory modules and policy networks on these behavioral traces, where the agent learns to predict or generate file operations that match a given user's behavioral profile; the agent's decisions are grounded in observed file-system state rather than free-form user input. The framework includes FileGramBench, a benchmark with synthetic personas and evaluation metrics (operation prediction accuracy, personalization fidelity, privacy preservation) to measure agent performance.
Production Impact
For teams building agentic systems that operate on local machines (coding assistants, data management tools, document processors), FileGram provides a practical path to personalization without requiring enterprise-wide trace collection or violating privacy regulations. Instead of building complex data pipelines to log user interactions across many organizations, you can use FileGramEngine to synthetically generate training data from a small number of representative persona definitions, reducing infrastructure complexity and privacy overhead. This enables faster iteration on personalization features: you can test agent behavior on diverse synthetic users without waiting for real-world data collection to accumulate. The main trade-off is synthetic-to-real gap—agents trained on simulated file-system traces may underfit to genuinely novel user behaviors or edge cases in production, requiring some fine-tuning on real traces or online adaptation mechanisms.
Limitations and When Not to Use This
FileGram assumes that file-system operations are sufficient to characterize user behavior, but this overlooks user intent, domain-specific context, or collaborative multi-user scenarios where behavior is harder to predict from traces alone. The synthetic data generation quality depends heavily on the accuracy of persona definitions—if your persona model misses important behavioral dimensions (e.g., security practices, rare workflows, or adversarial users), the trained agents will not generalize well to those populations. The paper does not address scalability to very large file systems (millions of files) or real-time personalization, where inference latency or memory consumption could become prohibitive. Additionally, the approach assumes privacy is preserved by using file-system traces, but metadata alone (file names, access patterns, timestamps) can still leak sensitive information in high-risk domains, requiring additional privacy-preserving techniques beyond the scope of this work.
Research Context
FileGram builds on two research streams: (1) agentic AI systems that operate in constrained environments (file systems, IDEs, desktops) and (2) personalization methods that learn user preferences from behavioral logs. The work extends prior work on synthetic data generation for training AI agents and responds to recent trends in local-first AI agents (e.g., GitHub Copilot, desktop automation agents) where privacy and decentralization are critical. It introduces a new evaluation benchmark (FileGramBench) for measuring agent personalization in file-system domains, parallel to benchmarks like GAIA or WebArena for grounded agent reasoning. The research opens a new direction: using file-system behavioral traces as a foundation for other downstream tasks (search, organization, security anomaly detection), potentially creating a richer foundation model for local-machine AI.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::