paper.tex

\documentclass[11pt]{article}

% Encoding
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}

% Professional preprint style: fonts, geometry, headings, title, header.
\usepackage{arxiv}

% Content packages
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage{listings}
\usepackage{subcaption}
\usepackage{enumitem}
\usepackage{natbib}

% hyperref loaded late, then coloured to match the palette.
\usepackage{hyperref}
\usepackage{url}
\hypersetup{
  colorlinks=true,
  linkcolor=arxivaccent,
  citecolor=arxivaccent,
  urlcolor=arxivaccent,
  filecolor=arxivaccent,
  breaklinks=true,
}

% Python code listing style
\definecolor{codebg}{rgb}{0.95,0.95,0.97}
\definecolor{codegreen}{rgb}{0.0,0.5,0.0}
\definecolor{codegray}{rgb}{0.5,0.5,0.5}
\definecolor{codepurple}{rgb}{0.58,0,0.82}
\definecolor{codeblue}{rgb}{0.0,0.0,0.7}

\lstdefinestyle{pythonstyle}{
    backgroundcolor=\color{codebg},
    basicstyle=\ttfamily\small,
    breakatwhitespace=false,
    breaklines=true,
    captionpos=b,
    commentstyle=\color{codegreen},
    keywordstyle=\color{codeblue}\bfseries,
    numberstyle=\tiny\color{codegray},
    stringstyle=\color{codepurple},
    language=Python,
    showstringspaces=false,
    numbers=left,
    numbersep=5pt,
    frame=single,
    rulecolor=\color{codegray},
    tabsize=4,
    morekeywords={dataclass, List, Optional, date, assert, Dict, field}
}

\lstdefinestyle{treestyle}{
    basicstyle=\ttfamily\small,
    frame=none,
    numbers=none,
    breaklines=false,
    backgroundcolor=\color{codebg},
}

% Style for conversation transcript excerpts (italic dialogue, framed).
\definecolor{convbg}{rgb}{0.98,0.97,0.92}
\definecolor{speakerblue}{rgb}{0.10,0.30,0.60}
\lstdefinestyle{conversationstyle}{
    backgroundcolor=\color{convbg},
    basicstyle=\ttfamily\footnotesize,
    breakatwhitespace=false,
    breaklines=true,
    captionpos=b,
    language={},          % plain transcript: no Python string/keyword colouring
    keywordstyle={},
    stringstyle={},
    commentstyle={},
    showstringspaces=false,
    numbers=none,
    frame=single,
    rulecolor=\color{codegray},
    tabsize=2,
}

% Style for runtime output / alert blocks (red border, fixed-width).
\definecolor{alertbg}{rgb}{1.00,0.96,0.94}
\definecolor{alertred}{rgb}{0.70,0.10,0.10}
\lstdefinestyle{alertstyle}{
    backgroundcolor=\color{alertbg},
    basicstyle=\ttfamily\footnotesize,
    breakatwhitespace=false,
    breaklines=true,
    captionpos=b,
    commentstyle=\color{codegreen},
    keywordstyle=\color{alertred}\bfseries,
    showstringspaces=false,
    numbers=none,
    frame=single,
    rulecolor=\color{alertred},
    tabsize=2,
    morekeywords={CRITICAL,WARNING,INFO,ALERT,DRUG-ALLERGY,CONFLICT},
}

\lstset{style=pythonstyle}

% Allow the one-column preprint to absorb long model names and citation groups
% without protruding into the margin.
\setlength{\emergencystretch}{3em}


\title{User as Code: Executable Memory for Personalized Agents}

\author{
  Bojie Li \\
  Pine AI
}

\date{}
\runningtitle{User as Code: Executable Memory for Personalized Agents}

\begin{document}
\maketitle

\begin{abstract}
An agent that remembers 100 trips should answer not only where a user traveled, but how often they visited Europe. Textual memories accessed through top-$k$ retrieval can recover individual trips but cannot reliably enumerate all of them, while conventional structured stores require a domain model to be defined and maintained. We introduce \textbf{User as Code} (UaC), which appends session facts and periodically materializes an LLM-generated typed Python view while retaining fact and archive channels for recall. On full 1{,}986-question LOCOMO suites, UaC records the highest observed same-backbone judge accuracy among six persistent-memory systems with GPT-5.6 Luna/Gemini~3 Flash Preview (78.8\%/80.6\% on 1{,}540 answer-bearing questions). Token F1 on those questions (39.4\%/46.9\%) and refusal accuracy on 446 adversarial questions (64.3\%/95.7\%) expose metric-specific tradeoffs. On 100 deterministically scored aggregate queries, both UaC and a raw-record Full Context+REPL reference reach 100\%, versus 43\% for MemMachine and 6\% for Mem0. The results show that an enumerable, code-readable representation enables reliable analytical queries while remaining compatible with retrieval-based factual memory.
\end{abstract}

\begin{center}
\small
Code: \url{https://github.com/19PINE-AI/user-as-code} \\[2pt]
Website: \url{https://01.me/research/user-as-code}
\end{center}
\vspace{-0.6em}

\input{body}

\bibliographystyle{plainnat}
\bibliography{reference}

\end{document}