Continuity gives AI agents a persistent memory. Every session, an agent wakes up blank — no memory of past work, no identity, no plan. Continuity solves this by externalizing the agent’s entire cognitive state into structured, versioned files that survive when the runtime dies.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.
Quickstart
Clone, copy templates, replace placeholders, and boot your agent in under 60 seconds.
Core Concepts
Understand loops, the three-class ontology, and why a single file is not enough.
State Packs
Pick a role-specific pack — developer, researcher, ops, orchestrator, and more.
File Reference
Full taxonomy of every state file, history directory, and library directory.
How Continuity Works
An AI agent’s context window is ephemeral and bounded. Continuity pushes everything that matters outside the window into structured files, then lets the agent selectively pull back only what it needs each session.Define
Declare who the agent is, what it knows, and what it is doing using structured markdown templates.
Track
The agent writes back to its state files after every meaningful action — tasks, learnings, snapshots.
The Three-Class Ontology
Every file in Continuity belongs to one of three classes:State
Root
.md files overwritten in place. Answers: what IS.History
Append-only nested directories. Answers: what WAS.
Library
Curated nested directories. Answers: what is KNOWN.
Pick Your Starting Point
Minimum Tier
8 files for simple agents: AGENTS, IDENTITY, SELF, SOUL, STATE, INIT, TODO, SNAPSHOT.
Standard Tier
16 files for most agents: adds USER, MEMORY, TOOLS, PLAN, CONTEXT, TASKS, LEARNINGS, HEARTBEAT.
Full Tier
27 files for operational agents with full runtime discipline.
Advanced Tier
Adds CHECKPOINT, AGENT_ACTIONS, and BINDING.schema.json for recursive multi-agent systems.
Runtime Adapters
Continuity is runtime-agnostic. Thin adapters wire it to any major agent framework:Claude Code
Entry via CLAUDE.md with auto-injection and writeback contract.
OpenAI Codex
Entry via AGENTS.md for Codex-based agent runtimes.
Hermes & OpenClaw
Adapters for Hermes, OpenClaw, and Agent Zero.
Generic Runtimes
Use Continuity with any agent that can read and write files.