Hermes Agent

Hermes Agent is model-agnostic, meaning you can point it at any OpenAI-compatible endpoint — Nous Portal, OpenRouter, local vLLM, or your own server. Its defining capability is a built-in learning loop: it writes reusable SKILL.md documents as it solves novel problems, so it compounds knowledge over time rather than starting fresh each session.

Hermes operates a three-level memory architecture: session memory (active conversation context), persistent memory (facts, preferences, and project details stored in ~/.hermes/ across all sessions), and hindsight memory (observations recalled from past sessions via fast FTS5 search). All data stays local on your machine — nothing is sent to any cloud.

Yes. Hermes supports Kanban-style multi-agent orchestration where a triage agent auto-decomposes a task into a tree of sub-tasks, assigns them to parallel workers, and synthesizes the results. The `delegate_task(background=true)` flag also allows background async sub-agents that return results as a new conversation turn without blocking your current session.

Hermes supports Telegram, Discord, Slack, WhatsApp (both Baileys bridge and official Meta Business Cloud API), Signal, Email, iMessage (via Photon), and the Raft agent network — all routed through a single gateway process. You can start a conversation on Telegram and continue it in your terminal with full cross-platform continuity.

Hermes has a dedicated MLOps mode for generating tool-calling trajectories in parallel with automatic checkpointing. It integrates with Atropos for reinforcement learning experiments, supports 11 tool-call parsers for training any model architecture, and exports conversations in ShareGPT format for fine-tuning. Trajectory compression also fits training data into tight token budgets.