Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/bitwikiorg/continuity/llms.txt

Use this file to discover all available pages before exploring further.

Continuity’s file-based state model is runtime-agnostic: any agent that can read and write markdown files can use it. This page covers three runtimes — Hermes, OpenClaw, and Agent Zero — each of which has a distinct model for loading agent instructions and injecting context. A generic four-step setup is included at the end for runtimes not covered here.

​
Hermes

Hermes is an agent runtime developed by NousResearch. It uses AGENTS.md as its primary instruction file and progressively discovers local context files as the agent moves through a project, making it a natural fit for Continuity’s scoped state chain model.

​
Entry file

AGENTS.md — Hermes loads this file from the project root. As the agent descends into subdirectories, it discovers additional context files and composes them into the active state.

​
Setup

Copy a Continuity pack into your project, then install the Continuity skill to the Hermes skills directory: 
cp -R skills/continuity ~/.hermes/skills/continuity
Prompt Hermes to use the skill: 
Use the continuity skill. Boot this project. If READY, continue with the top task in TODO.md.
Hermes can read and write files directly. After each session, update STATE.md, TODO.md, and SNAPSHOT.md in place.

​
Auto-injected instance files

Hermes auto-injects three files from the Hermes home path (~/.hermes/) into every session — independently of the project-level state chain:
FileWhat it holds
SOUL.mdAgent essence, purpose, and constitution
MEMORY.mdDurable facts index
USER.mdHuman model — who the agent is serving
Copy these files from your Continuity workspace to ~/.hermes/ to ensure they are auto-loaded on every Hermes session. These are instance-level state files — they persist across projects, unlike project-level files such as STATE.md or TODO.md.

​
Progressive discovery and AGENTS.promptinclude.md

Hermes supports AGENTS.promptinclude.md for progressive discovery. Files listed there are injected into the agent’s context on every session without manual loading. Use this to wire in your required-tier state files: 
# AGENTS.promptinclude.md

@AGENTS.md
@STATE.md
@TODO.md
@SNAPSHOT.md

​
Security scanning

Hermes supports security scanning during context loading. Use CHECKPOINT.md to define approval gates for high-risk or irreversible actions. WARNING.md carries the hard limits and pre-action checklists that Hermes evaluates before executing tasks.

​
OpenClaw

OpenClaw uses SKILL.md directories with explicit loading order, allowlists, and security gates. Its workspace template model — BOOT, BOOTSTRAP, SOUL, IDENTITY, HEARTBEAT, TOOLS, and USER — directly inspired Continuity’s boot-and-identity file split.

​
Entry file

AGENTS.md — OpenClaw uses AGENTS.md as the primary instruction file, composed via the skill’s YAML config wrapper.

​
Setup

Copy a Continuity pack into your project and install the skill: 
cp -R skills/continuity ./.agents/skills/continuity
Create a YAML config wrapper that declares the entry file, boot file, tracked state files, writeback targets, and security rules: 
# continuity-agent.yaml
entry_file: AGENTS.md
boot_file: INIT.md
state_files:
  - STATE.md
  - TODO.md
  - SNAPSHOT.md
writeback:
  events: logs/events.jsonl
  snapshots: snapshots/
rules:
  - do_not_write_secrets_to_markdown
  - require_checkpoint_before_irreversible_actions
Prompt OpenClaw: 
Boot project and continue next valid action.

​
Boot and identity file split

OpenClaw’s native workspace template maps cleanly onto Continuity’s file clusters:
OpenClaw fileContinuity equivalentCluster
BOOTBOOT.md + INIT.mdBoot
BOOTSTRAPBOOTSTRAP.mdBoot
SOULSOUL.mdIdentity
IDENTITYIDENTITY.mdIdentity
HEARTBEATHEARTBEAT.mdState
TOOLSTOOLS.mdCapability
USERUSER.mdIdentity
If you are already using OpenClaw’s native file structure, you can map your existing files onto Continuity templates without renaming — just add the missing State, Memory, and Safety cluster files.

​
Security gates

OpenClaw’s security gate system integrates directly with Continuity’s safety cluster. Declare high-risk gates in CHECKPOINT.md and reference them from your YAML config’s rules block. OpenClaw evaluates these gates before executing any action flagged as irreversible.

​
Agent Zero

Agent Zero does not auto-load AGENTS.md the way Codex or Hermes do. Instead, it uses a prompt-include system: files listed in AGENTS.promptinclude.md are injected into the agent’s system prompt automatically on every session.

​
Entry file

AGENTS.md — but loaded via AGENTS.promptinclude.md rather than auto-discovery. AGENTS.promptinclude.md is the effective trigger file for Agent Zero.

​
Setup

Copy a Continuity pack into your project, then create AGENTS.promptinclude.md to wire in your state files: 
# AGENTS.promptinclude.md

@AGENTS.md
@STATE.md
@TODO.md
@SNAPSHOT.md
Prompt Agent Zero: 
Read STATE.md and TODO.md. If READY, execute the next valid action.
After work, update continuity files.
Agent Zero can read and write files via its built-in tools. After each session, update STATE.md, TODO.md, and SNAPSHOT.md in place.

​
Multi-agent subordinates

For multi-agent work with Agent Zero, use the orchestrator pack. Each subordinate agent gets scoped access to its portion of the state chain. Use BINDING.schema.json to validate agent contracts before spawning subordinates — it defines who runs the computation, what tools are allowed, what the budget is, and when to stop.
When using Agent Zero’s subordinate system, nest state files per agent scope. Each subordinate gets its own STATE.md and TODO.md under its working directory. The root orchestrator’s AGENTS.md governs the overall load order; subordinate-scoped files refine it without weakening parent safety rules.

​
Generic / other runtimes

Any agent that can read markdown files can use Continuity. If your runtime is not listed above, follow these four steps:
1

Copy needed templates to your workspace

Copy the templates matching your tier from the Continuity repo. The Minimum tier is enough for most agents:
cp continuity/{AGENTS,IDENTITY,SELF,SOUL,STATE,INIT,TODO,SNAPSHOT}.md /your/workspace/
Fill in all {{UPPER_CASE}} placeholders before booting.
2

Read AGENTS.md, then INIT.md, then files in load order

Point your runtime at AGENTS.md as the entry file. On first boot, run the INIT.md nine-stage reconstitution procedure. After that, follow the load order declared in AGENTS.md at the start of every session.
3

Remove AGENTS.promptinclude.md if not needed

AGENTS.promptinclude.md is only needed for runtimes that support auto-injection (Agent Zero, Cursor, Hermes). If your runtime handles file loading natively or you reference state files explicitly in your entry file, remove AGENTS.promptinclude.md to keep the workspace clean.
4

Implement writeback after each session

After every meaningful action, update the files whose claims are no longer accurate:
  • STATE.md — new operational truth
  • TODO.md — task queue with completed items moved to completed/
  • SNAPSHOT.md — refreshed restoration anchor
  • LEARNINGS.md — if new insights emerged
Without writeback, Continuity degrades to static config. The feedback loop — read state, act, update state — is what makes it accumulate value across sessions.
The system is runtime-agnostic by design. The templates are plain markdown with {{VARIABLES}} placeholders. Any agent that reads files can boot from them; any agent that writes files can update them. The only hard requirement is that the agent follows its own specification — reads the state chain before acting and writes it back when reality changes.

Build docs developers (and LLMs) love

Get started for freeTalk to us