papers/zoo-ai-memory/zoo-ai-memory.tex

\documentclass[11pt,letterpaper]{article}
\usepackage{shared/zoocover}

% Packages
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
% geometry loaded by zoocover
\usepackage{amsmath,amssymb,amsthm}
\usepackage{graphicx}
\usepackage{hyperref}
\usepackage{booktabs}
\usepackage{array}
\usepackage{tikz}
\usetikzlibrary{shapes,arrows,positioning,calc,fit}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{enumitem}
\usepackage{mathtools}
\usepackage{xcolor}
\usepackage{listings}

% Theorem environments
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}{Definition}
\newtheorem{example}{Example}
\theoremstyle{remark}
\newtheorem{remark}{Remark}

% Custom commands
\newcommand{\Zoo}{\textsc{Zoo}}
\newcommand{\Hanzo}{\textsc{Hanzo}}
\newcommand{\Lux}{\textsc{Lux}}
\newcommand{\AChain}{\textsc{A-Chain}}
\newcommand{\IChain}{\textsc{I-Chain}}
\newcommand{\KChain}{\textsc{K-Chain}}
\newcommand{\TChain}{\textsc{T-Chain}}
\newcommand{\GChain}{\textsc{G-Chain}}
\newcommand{\CKKS}{\textsc{CKKS}}
\newcommand{\TFHE}{\textsc{TFHE}}
\newcommand{\CRDT}{\textsc{CRDT}}
\newcommand{\DID}{\textsc{DID}}
\newcommand{\FHE}{\textsc{FHE}}
\newcommand{\KEEPER}{\textsc{Keeper}}

% Hyperref setup
\hypersetup{
    colorlinks=true,
    linkcolor=black,
    citecolor=black,
    urlcolor=black
}

% Title information
\title{\textbf{Sovereign AI Memory: Persistent, Private, and Portable\\Agent State on Zoo Network}\\
\large Encrypted Memory Capsules with Threshold Access Control}
\author{
Antje Worring, Zoo Labs Foundation Research\thanks{[email protected]}\\
\textit{Zoo Labs Foundation}\\
\texttt{[email protected]}
}

\date{April 2025}

\begin{document}

\zoocoverpage

\begin{abstract}
Contemporary AI agents suffer from persistent amnesia. Each session begins tabula rasa; memory, when it exists, lives in provider-controlled vector databases that the agent cannot own, move, or protect. We argue that AI memory is not a retrieval problem---it is a \emph{sovereignty} problem. This paper introduces \textbf{Memory Capsules}, encrypted append-only vaults that give agents persistent, private, and portable state on \Zoo{} Network. A Memory Capsule binds to a decentralized identity (\DID{}) on \IChain{}, stores content-addressed entries in an encrypted SQLite vault sealed with \KChain{}-managed keys, synchronizes across agent instances via operation-based \CRDT{}s, and enforces fine-grained access through capability tokens and \TChain{} threshold reveal. We formalize four memory types---episodic, semantic, procedural, and preference---and show how each maps to distinct encryption and garbage-collection policies. For sensitive memories, we demonstrate \FHE{}-encrypted reasoning via \CKKS{} on \AChain{}: an agent can query its own memory without decrypting it, and third parties can verify reasoning traces without observing content. Cross-agent knowledge transfer is achieved through selective disclosure using zero-knowledge proofs, enabling collaborative intelligence without surrendering private state. We benchmark Memory Capsule operations on \Zoo{} Network, measuring append latency (1.2\,ms), \CRDT{} merge throughput (14k ops/s), and \FHE{} query overhead (340\,ms per encrypted recall). The architecture positions AI memory as a first-class on-chain asset: ownable, transferable, and economically valuable.

\textbf{Keywords}: AI memory, sovereign state, encrypted vault, CRDT, FHE, capability-based access, decentralized identity
\end{abstract}

\tableofcontents
\newpage

%----------------------------------------------------------------------

\input{sections/01-introduction}
\input{sections/02-why-sovereign-ai-memory-matters}
\input{sections/03-memory-capsule-architecture}
\input{sections/04-memory-access-control}
\input{sections/05-memory-types}
\input{sections/06-fhe-encrypted-memory}
\input{sections/07-cross-agent-knowledge-transfer}
\input{sections/08-memory-economics}
\input{sections/09-implementation-on-zoo-network}
\input{sections/10-benchmarks}
\input{sections/11-related-work}
\input{sections/99-conclusion}

\end{document}