2026-04-14 concept agents tools memory skills open-source

Hermes Agent

Self-improving AI agent from Nous Research. 163.1k stars, 26.7k forks, MIT (up from 76.6k stars in April 2026 – 2x growth in one month, top-tier traction for an agent harness). The agent creates skills from experience, improves them during use, and maintains memory across sessions.

Not a framework for building agents. This is a ready-to-run agent with learning built in.

What makes it different

Learning loop: agent creates skills autonomously, then self-improves them at runtime. Skills are compatible with agentskills.io open standard.

Cross-session memory: FTS5 session search + LLM summarization. Plus “Honcho dialectic user modeling” that builds a profile of the user over time.

Multi-platform native: one gateway process handles Telegram, Discord, Slack, WhatsApp, Signal, Email. Same conversation continues across platforms. Voice memos transcribed automatically.

Seven deployment backends:

Local, Docker, SSH (standard)
Daytona, Modal, Vercel Sandbox (serverless with hibernation, near-zero cost when idle)
Singularity (HPC containers)

200+ models: any OpenAI-compatible API. First-class wiring for Nous Portal, OpenRouter, NovitaAI, NVIDIA NIM (Nemotron), Xiaomi MiMo, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, plus self-hosted endpoints. Switch with hermes model, no code changes.

Native Windows beta: PowerShell installer ships portable MinGit (~45MB), isolated from any system Git. Termux on Android works via curated extras. WSL2 remains the most battle-tested Windows path.

Architecture

CLI TUI / Telegram / Discord / Slack / WhatsApp / Signal / Email
    ↓
  Gateway (single process, all platforms)
    ↓
  Agent Loop (tool calls, responses, learning)
    ↓
  Terminal Backend (local / Docker / SSH / Daytona / Modal / Singularity / Vercel Sandbox)

40+ built-in tools, MCP server integration, parallel subagents, Python RPC, cron scheduler.

Research angle

Nous uses Hermes for training data: batch trajectory generation, trajectory compression, Atropos RL environments. The agent’s tool-calling traces become training data for the next generation of models.

This connects to decision-traces-compound: every agent action becomes a training signal. And to asi-evolve-ai-research-agents: both systems use agent outputs to improve the underlying models.

How it compares to our stack

	Hermes Agent	Our approach (sgr-agent + Claude Code)
Language	Python	Rust
Learning	Auto-creates skills	Skills authored in SKILL.md, hot-reloaded
Memory	FTS5 + LLM summarization	Solograph (FalkorDB + vectors)
Deployment	7 backends including 3 serverless (Daytona, Modal, Vercel Sandbox)	Local + competition RPC
Models	200+ via OpenRouter	Provider-agnostic (config.toml)
Tools	40+ Python	14 generic + domain middleware

Key insight from Hermes: skill self-improvement at runtime is something our self-evolution pattern describes in theory (fixed/mutable partition). Hermes implements it in production.

The serverless backends (Daytona, Modal) are also interesting for infra-two-tools: agent hibernates when idle, wakes on demand.

Growth signal (2026-05-23)

From 76.6k → 163.1k stars in 39 days (~85k new stars, ~2.2k/day). Forks crossed 26.7k. For comparison, ruflo-orchestration and bmad-method sit two orders of magnitude lower. This is the most-starred ready-to-run agent today and the strongest empirical signal that the agentic harness layer – not the model layer – is where developer attention concentrates. Matches the thesis in claude-code-anatomy: harness eats the model.