The Diplomat Didn't Build a Chatbot
Singapore’s Foreign Minister published his personal AI architecture last week.
It runs on a Raspberry Pi. It processes his WhatsApp messages and voice notes. It drafts speeches, compiles briefings, and answers questions about diplomatic history — drawn from a growing graph of everything he has read, said, and decided. It never leaves his private network.
His name is Vivian Balakrishnan. He has managed Singapore’s foreign policy for over a decade. He is also a trained ophthalmologist who codes.
He calls the system NanoClaw.
The coverage has focused on the novelty of it. A cabinet minister building his own AI. Running on a $90 computer. Using Claude.
That’s the wrong story.
The architectural choice nobody mentioned
Here’s what’s structurally interesting about NanoClaw.
It doesn’t do RAG.
Most AI systems that work with personal or organizational knowledge use Retrieval-Augmented Generation: you store documents, chunk them, embed them, and retrieve the most relevant chunks when the model needs context. It works. It’s fast. It’s also fundamentally a retrieval tool — it gives the model access to raw material, not structured understanding.
Balakrishnan built something different. A system called mnemon sits at the core of NanoClaw. It doesn’t store document chunks. It stores synthesized facts — discrete, structured nodes in a knowledge graph, organized by entity, concept, and timeline. When he asks a question, the system runs a semantic query against synthesized knowledge, not raw text. The model answers from structured understanding, not from retrieved fragments.
That’s not a technical detail. That’s the difference between a search interface and a memory system.
RAG answers questions about documents. A knowledge graph answers questions about what you know.
The distinction matters because most of what gets lost in modern knowledge work isn’t documents. It’s the synthesis that exists between documents. The decision that was made across three Slack threads. The commitment that appears in a meeting summary but not in any tracker. The context that was clear in the moment and is now scattered across six tools.
Chunking and retrieving those tools doesn’t recover the synthesis. It just gives you the raw material again. You’re still doing the work of connecting it.
Balakrishnan’s system does the connecting.
The same week, at a different scale
On April 22, Google Cloud Next in Las Vegas drew a different crowd. Enterprise CTOs. Platform engineers. The companies running production AI at scale.
The headline from analysts who covered all three days: the axis of competition in AI has shifted.
Not shifted from one model to another. Shifted from intelligence to governability.
The average enterprise now runs 12 AI agents. Eighty-nine percent of business teams are already using them. A2A protocol — Google’s Agent2Agent standard — is in production at 150 organizations. The announcements weren’t about smarter models. They were about agents that can coordinate with each other without needing human intermediaries to relay context between them.
The problem is not that the agents are too dumb. The problem is that they don’t know what the other agents decided.
The problem is the same at both scales. A foreign minister with one AI system, and a Fortune 500 running twelve. The intelligence is available. The coordination isn’t.
The piece everyone is missing is above the models.
Why most companies are still doing it wrong
The instinct when deploying AI is to think about capability. What can the model do? What tools can I connect? How do I get it to answer questions about my knowledge base?
That’s the RAG mental model. Capability + retrieval. Input your documents, ask your questions, get your answers.
It works for isolated queries. It doesn’t work for the real problem.
The real problem isn’t answering questions about documents. The real problem is that knowledge work generates a continuous stream of open loops — decisions made without closure, commitments made without tracking, context generated without synthesis — and no existing system manages that stream. The documents pile up. The Slack threads multiply. The meeting summaries accumulate. The model gets smarter. The overhead grows.
Balakrishnan’s architecture sidesteps that problem by treating information as something to synthesize, not retrieve. Every incoming signal — a speech, a news article, a voice note, a WhatsApp conversation — gets processed into structured facts that join the knowledge graph. The system doesn’t just store what happened. It builds a growing understanding of what it means.
That’s not how most companies are building their AI infrastructure. Most companies are still pointing models at document stores and calling it a second brain.
The reframe
The Singapore FM story isn’t about what’s possible when you’re a diplomat who codes. It’s about what architecture actually works for personal AI infrastructure — and why the dominant pattern (RAG + chat interface) is insufficient for the use case that matters most.
The Google Cloud Next story isn’t about platform competition. It’s about market confirmation. When the biggest cloud provider’s flagship event shifts from “our models are smarter” to “our coordination layer is more governable,” the intelligence-as-competitive-advantage era is functionally over. The race moved.
Taken together: the people who actually have something at stake — a minister managing state relationships, enterprises running production operations — are building coordination infrastructure. Persistent memory. Synthesized knowledge. Systems that close loops, not just answer questions.
The question is not whether AI will be useful for your organization.
The question is whether you’re building something that synthesizes and persists, or something that retrieves and forgets.
One of those is a chat tool with a longer context window.
The other is infrastructure.
Eliran Keren — Founder of Deeplica, building the coordination layer for humans who’d rather direct than operate.
Sources: Singapore FM Builds NanoClaw AI Second Brain — Awesome Agents · Designing Memory for AI Agents: Inside LinkedIn’s Cognitive Memory Agent — InfoQ · Google Cloud Next 2026: A Structural Analysis — DEV Community · Google Cloud Next 2026: AI Agents, A2A Protocol — The Next Web
