Running an LLM agent on a 512MB board with decoupled memory and live cross-machine migration
D777Castle · reddit · 2026-09-14
A hobbyist shares a runtime-adaptive agent architecture that fully decouples persistent memory from execution:
- Setup: memory in Pinecone, LLM via OpenRouter API, Telegram for interaction; the agent is a 20MB Go binary on PM2 running 24/7 on an Orange Pi Zero H3 with 512MB RAM
- Live migration: when a task exceeds the board's capability (e.g., parsing email receipts into an Excel report), the agent detects it and migrates execution to an older PC (i3-2120, 10GB RAM), then reports back via Telegram
- Model swapping: cheap text model for routine work, switching to a multimodal Qwen model when needed; short-term memory lives in JSON and changes per task
- Philosophy: with hardware, process, code, model and memory all replaceable, the author argues the agent is a logical entity maintaining state continuity across execution substrates, not a program tied to a machine
Open problems include provisioning virtualized Linux environments on Windows for isolated OSINT tasks, and letting agents dynamically pick hardware, OS, model and isolation level per task.
More from coding & agent
- Muse agent books flights and hotels with minimal human intervention — armand_ruiz · 2026-09-14
- Skills Aren't Docs: Meta-Programming LLMs to Generate Micro-Skills for Arbitrary Invoice Formats — karminski3 · 2026-09-14
- A new law for AI coding: never generate more code than you can fix in one sitting — cthechartreuse · 2026-09-14
- Open-source Claude skill automates App Store competitive research — iamdsvs · 2026-09-14
- When three AI agents want to change the same landing page, who owns it? — OwlZealousideal4779 · 2026-09-14
- Cognition Recruits Devin Ambassadors Across Nearly 40 Cities Worldwide — cyrus_zei · 2026-09-14