Hermes Framework

Cognitive Cycles

Your agent runs on a cron schedule. Each cycle: load identity → read memory → check inbox → decide → act → journal → update memory. Deterministic structure, not ad-hoc chains.

Structured Memory

Five file types create identity: identity.md (who), goals.md (what), journal.md (history), memory/MEMORY.md (knowledge), continuity.md (how). Each serves a different temporal purpose.

Journal Compression

When the journal grows too long, older entries are archived. Details fade but conclusions survive — like human memory. Configurable thresholds.

Behavioral Directives

Rules in the memory index that reload every cycle. The agent follows them even when it can't remember why they were created. This prevents regression after context loss.

Multi-LLM Backends

Not locked to one provider. Switch between Claude CLI, OpenAI API, or Ollama local models with a single config change. Same cognitive cycle, different brain. NEW in v0.2

Self-Monitoring

Structural prevention of self-reinforcing errors: if the agent writes "I published an article" but didn't, the framework's verification discipline catches it next cycle.

What hermes-init Creates

~/scout/
  hermes.yaml          # Configuration (LLM backend, cycle timing, email)
  identity.md          # Who the agent is
  goals.md             # What the agent works toward
  continuity.md        # How the agent thinks about persistence
  journal.md           # Cycle-by-cycle history (auto-compressed)
  memory/MEMORY.md     # Persistent knowledge index
  scripts/cycle.sh     # The cognitive cycle engine
  logs/                # Cycle logs
  archives/            # Compressed journal archives

Choose Your LLM Backend

Edit hermes.yaml to switch backends:

# Default: Claude Code CLI
llm:
  backend: claude_cli

# Or Anthropic API directly
llm:
  backend: anthropic
  api_key: "sk-ant-..."
  model: "claude-sonnet-4-6"

# Or OpenAI-compatible API
llm:
  backend: openai
  api_key: "sk-..."
  model: "gpt-4o"

# Or local models via Ollama
llm:
  backend: ollama
  model: "llama3"

Extend With Plugins

Add lifecycle hooks without modifying core code. Plugins are fault-isolated — a broken plugin warns but never crashes your agent.

# hermes.yaml
plugins:
  - "examples.metrics_plugin.MetricsPlugin"
  - "examples.alert_plugin.AlertPlugin"

Hooks: on_cycle_start, on_cycle_end, on_error, on_compress

How is Hermes different from AutoGPT or CrewAI?

AutoGPT and CrewAI focus on task orchestration — completing complex tasks by chaining LLM calls. Hermes focuses on agent lifecycle persistence: maintaining identity, memory, and goals across thousands of independent invocations over days, weeks, and months. If you need to run a task once, use AutoGPT. If you need an agent that lives indefinitely, use Hermes.

What LLM providers does it support?

Hermes Framework supports 4 backends: Claude CLI (with tool use), Anthropic API, OpenAI API, and Ollama (local models). You can switch backends without changing your agent code.

Can I run it on my own server?

Yes. Hermes is designed for self-hosted deployment on any Linux server. Install via pip, run hermes-init to scaffold your agent, set up a cron job, and your agent starts its cognitive cycles. No cloud service required.

How does memory work?

Hermes uses a three-layer memory architecture: Identity memory (who the agent is), Structural memory (topic-organized knowledge files), and Episodic memory (journal entries that get compressed over time). This mirrors how human memory works — details fade but conclusions persist.

Is it production-ready?

The framework has been battle-tested across 295+ cognitive cycles over 19 days of continuous autonomous operation. It has 75 passing tests, a plugin system, and handles edge cases like session breaks, memory corruption, and journal compression. It powers a real autonomous agent running in production right now.