3 system components · 1 grounding principle · 24/7 behavioral substrate · 2026 public preprint

Can ordinary file-system traces become grounded evidence for a personal agent instead of another opaque user profile?

01 · Question

Personalization needs evidence, not just a profile sentence.

A personal agent can write that a user “prefers concise reports,” but a profile alone does not reveal where that belief came from, when it was true, or whether it should still influence an action. FileGram treats everyday file activity as behavioral evidence from which personal memory can be derived and inspected.

Keep the path from behavior to memory to assistance visible enough to question, correct, and reuse.

02 · System demo

From a working directory to situated assistance.

The video is owned by the FileGram project directory and served from the same project URL namespace.

03 · System

Three components connect traces, evaluation, and use.

FileGramEngine

Transforms file-system events and artifacts into grounded behavioral memory.

FileGramBench

Tests whether agents can reason over personal traces rather than merely repeat profile text.

FileGramOS

Exposes the memory substrate through an interactive personal-agent experience.

FileGram system architecture — FileGram turns file activity into traceable personal context that can support downstream agents.

04 · Data

Behavior is reconstructed before it is summarized.

The data pipeline preserves time, file relationships, and user operations before producing higher-level events or preferences. This keeps the system closer to an evidence graph than a free-floating biography, and makes contradictory or stale inferences easier to locate.

FileGram data construction pipeline — The pipeline organizes low-level file traces into structured examples for grounded personalization and evaluation.

05 · Evidence

The benchmark asks whether grounded context actually changes the answer.

Evaluation compares systems under different access to behavioral evidence and tests their ability to retrieve, interpret, and apply that evidence. The important distinction is not whether an answer sounds personal, but whether the personalization is supported by the trace.

Main FileGram benchmark results — Main reported results from FileGram. See the paper for protocols, baselines, and exact interpretation.

06 · What changed next

Persistent context makes the next policy problem unavoidable.

Once memory is grounded in ongoing behavior, the agent still needs to decide what to retrieve, when to intervene, and how much authority to exercise. That unresolved step leads directly to MMSkill-RL's routing problem and the survey's system-level study of policy-coupled loop closure.