Second Self: A Memory-Augmented Autonomous Digital Twin Using LLM Agents and Structured Personal Memory

Routine digital work frequently requires switching between email clients, calendars, browser tabs, and team collaboration tools. While a personal assistant that retains user preferences can reduce repetitive contextsetting, persistent memory and cross-application actions also introduce privacy and operational control challenges. This paper presents Second Self, a memoryaugmented digital twin prototype that combines structured personal-memory files with a language-model planning workflow and modular desktop and browser interfaces. Our system architecture integrates a macOSoriented desktop client, a Python asynchronous orchestrator, a companion Next.js and FastAPI web application, and an onboarding pipeline that constructs readable identity and voice-profile files. In the current implementation, the orchestrator executes multi-step plans through Anthropic-format tool calls alongside a modular Model Context Protocol (MCP) tool factory, while connecting to configured external services including hosted language models, Google Workspace APIs, web search, and cloud authentication. We therefore position Second Self as a local-first, memoryaugmented prototype that balances structured local context with cloud-backed reasoning. To validate our core software modules in a cross-platform test environment, we executed a verification suite covering 530 unit test definitions, passing 87 portable mocked component tests and four synthetic Gmail authentication adapter checks. We also detail the design of our offline evaluation harness and outline the controlled humansubject and end-to-end macOS benchmarks planned for future work.