Singapore FM demos personal AI agent

Singapore Foreign Minister Vivian Balakrishnan used a May 16 appearance at AI Engineer Singapore to make a case for hands-on use of artificial intelligence by public officials. In a speech published by Singapore’s Ministry of Foreign Affairs, Balakrishnan said people in authority could delegate work but not accountability, and described building tools himself as part of understanding the technology he discusses publicly. A GitHub gist published from Balakrishnan’s account about three weeks earlier laid out the system he called “NanoClaw — Personal Claude Assistant (second brain for a diplomat).” The document said the assistant runs on a Raspberry Pi, uses Anthropic’s Claude, connects to WhatsApp, Telegram, Slack and Discord, and stores information over time through a structured memory system rather than a standard chat history. (mfa.gov.sg) ### What exactly did Balakrishnan show? Balakrishnan’s GitHub materials described a personal assistant designed to answer questions, process voice and images, schedule recurring tasks and pull in context from prior documents and conversations. The gist said the system was built as a “self-hosted, compounding-memory AI assistant” and framed it as a tool for diplomatic work. (gist.github.com) The Raspberry Pi setup was paired with messaging channels rather than a standalone web app. The repository linked to Balakrishnan’s account said users could add channels including WhatsApp, Telegram, Discord, Slack and Gmail, with separate group contexts and containerized environments. ### Why did he connect the demo to governing AI? At the May 16 event, Balakrishnan said, “The one thing which you cannot outsource is your personal understanding.” He added that a person “in a position of authority” could delegate work but “cannot delegate accountability,” language that tied the demo to how decision-makers should approach AI systems they regulate or deploy. (gist.github.com) (github.com) The ministry transcript did not use the exact line circulating in social posts about being “briefed on,” but it did show Balakrishnan arguing that direct use matters more than second-hand descriptions. He told the audience that “the tools have already been invented” and that the task now is getting people to understand and assemble them. (mfa.gov.sg) ### What is inside the system he published? The GitHub gist said the assistant uses three layers: raw source files, a knowledge graph called “mnemon,” and synthesized wiki pages. The raw inputs include speech transcripts, articles and web clips, while the graph stores discrete facts with timestamps, tags and links to related entries in a SQLite-backed database. The same document said the recall process runs a semantic query every time the agent is invoked and injects relevant entries before the model responds. (mfa.gov.sg) Balakrishnan’s public GitHub account also shows a fork of the “mnemon” memory project, which he describes as “LLM-supervised persistent memory for AI agents.” ### Where does Anthropic fit in? Anthropic’s Claude is the language model named in Balakrishnan’s gist, and NanoClaw is presented there as a Claude-based assistant rather than a model built from scratch. (gist.github.com) The gist says the system runs on Anthropic’s model while the Raspberry Pi handles orchestration, routing and local storage. NanoClaw itself is also available as a public GitHub project. (gist.github.com) Its repository describes it as a lightweight alternative that runs in containers, connects to messaging apps and uses Anthropic’s Agents SDK. ### What comes next for readers who want to verify the demo? The Ministry of Foreign Affairs has published Balakrishnan’s full May 16 speech on its website, and his GitHub gist and related repositories remain public as of May 17. (gist.github.com) Those materials name the components he used — Raspberry Pi, Claude, mnemon and messaging integrations — and provide the clearest primary-source record of the build he discussed. (mfa.gov.sg) (github.com)