Concepts
Core Concepts & Definitions
Understanding OpenJellyfish starts with these terms. Each one answers the same question: why intelligence should live in documents.
Glossary
- Document-first Agent
- In a document-first agent, the agent's identity, memory, skills, workflow, permissions, and knowledge all exist as documents. The runtime only reasons and executes — it carries no business logic or persona. The same runtime, paired with different documents, becomes an entirely different agent.
- Document-native Intelligence
- Document-native intelligence means the prompt is just one projection of the documents. Memory, workflow, skills, policies, permissions, and knowledge are all documents too. Every customizable capability comes from documents rather than database rows or hard-coded logic.
- Persistent Intelligence
- Because an agent is just a set of documents, it can persist independently of any specific runtime. The harness can be upgraded, the LLM swapped, the runtime rewritten — yet the agent's soul remains in its documents and never disappears with them.
- Agent Harness
- The harness is not the agent — it is the body that lets intelligence run. It does not know who it is, what it should do, or what memory and permissions it has. It stays simple forever. Keep complexity in the documents, and one harness can host unlimited, entirely different agents.
- Environment Projection
- The environment includes the filesystem, tools, MCP, scheduler, and external services. It does not belong to the agent — it is the world the agent lives in. The harness interacts with the environment, and documents define how the harness understands and projects onto it.
- Document Memory
- Memory is not stuffed into an opaque vector store or database — it is written as readable, browsable, version-controlled documents. You can walk through what an agent remembers layer by layer, like a memory palace, and edit it directly.
An agent whose intelligence lives in documents, not in prompts or orchestration code.
Intelligence defined, stored, and evolved with documents as its first-class medium.
Intelligence that outlives any runtime, database, or platform and can be migrated freely.
The minimal body of intelligence that only does Observe · Think · Decide · Act.
How intelligence projects itself onto the world via filesystem, tools, and services.
Long-term, layered memory that exists as readable, editable documents.
Frequently Asked Questions
What is OpenJellyfish?+
OpenJellyfish is a document-first AI agent framework. Instead of writing intelligence into prompts or orchestration code, it keeps intelligence in documents: the LLM provides intelligence, the harness runs it, documents define it, and the environment projects it. Open source, self-hosted, and git-portable.
What is a document-first agent?+
A document-first agent stores the agent's identity, memory, skills, workflow, permissions, and knowledge entirely in documents, while the runtime only reasons and executes. Swap the documents and you swap the agent — the runtime stays the same.
Why documents instead of prompts?+
A prompt is just one projection of the documents, and it is hard to version, copy, and migrate. Putting intelligence in documents gives you a readable, diff-able, merge-able, copyable lifecycle — the entire agent can be maintained with mature software engineering tools.
How does the harness work?+
The harness is the minimal body of intelligence, running only an Observe · Think · Decide · Act loop. It carries no business logic or persona; all customization happens in the documents, so one harness can run countless different agents.
What is the environment?+
The environment is the world the agent lives in: filesystem, tools, MCP, scheduler, and external services. The agent reaches the real world and manifests results through the environment, but the environment itself is not part of the agent.