Nous Research went from fine-tuning Hermes LLMs to owning the full inference-to-action stack — 8.7k GitHub stars, 142 contributors, MIT licensed, model-agnostic across OpenRouter's 200+ endpoints. But let's be precise about "self-improving": the agent auto-generates reusable skill files after 5+ tool calls and stores 3,575 characters of prompt memory (~1,300 tokens total). No model weights change, ever — it's procedural memory and workflow caching dressed up in AGI-adjacent language. The multi-platform gateway running Telegram/Discord/Slack/Signal from a single process with cross-session continuity is where this actually gets sticky, because distribution and persistent state across chat surfaces is the unsexy moat that compounds while everyone else debates which foundation model is smartest.

Top comment by @Benthic