One config. Every AI agent. Persistent memory.

AgentFS turns your Obsidian vault into a unified operating system for AI agents. Define your identity, memory, and security rules once — compile to native formats for Claude Code, Cursor, OpenClaw, and others.

AI agents today are silos. Claude Code needs CLAUDE.md, Cursor needs .cursor/rules/, and OpenClaw needs .openclaw/. You maintain the same rules, identity, and context in three different places. Agents don't share memory. They don't respect the same security policies. They start from zero every session.

Most importantly: Agents lack a "Persistent State." They can't "pause" and "resume" tasks across sessions or hand off context to other agents without major manual intervention.

AgentFS introduces a Kernel Space (.agentos/) — a single source of truth for who you are, what you know, and what's off-limits. One command compiles it into native configs for every agent you use.

• Persistent State: Agents can maintain state across toolchains and sessions.
• Inter-Agent Sync: A shared filesystem where different agents exchange context without API overhead.
• Secure Context Isolation: Deterministic security rules that agents actually respect across environments.

AgentFS isn't just a config generator; it's the missing infrastructure layer for autonomous AI agents. While others build the agents, AgentFS builds the environment they live in.

.agentos/                        CLAUDE.md
  manifest.yaml     agentfs      .claude/settings.json
  init.d/          -------->     .cursor/rules/*.mdc
  memory/           compile      .openclaw/
  security/                      AGENT-MAP.md

# Install globally from npm
npm install -g create-agentfs

# Verify
agentfs --version

Or use without installing:

npx create-agentfs my-vault

# Update to latest version
npm install -g create-agentfs@latest

# Upgrade existing vault to new version
agentfs upgrade --dir ~/my-vault

npm uninstall -g create-agentfs

agentfs init my-vault
cd my-vault
agentfs compile

That's it. Your vault now has a kernel, and every supported agent gets native configuration.

For detailed setup with interactive onboarding, modules, and profiles — see the Quick Start Guide.

Long-lived agents like Cowork, OpenClaw, or Cursor in background mode get the complete experience: memory accumulates between sessions, cron jobs consolidate knowledge, the agent "grows" over time.

agentfs compile                   # compile kernel → native configs
agentfs memory consolidate        # snapshot session memory
agentfs sync                      # sync memory back to kernel

The agent reads semantic.md on boot, writes episodic memories, runs consolidation at session end. Full read-write cycle.

Short-lived sessions (Claude Code in terminal, one-off Cursor tasks) use AgentFS as a smart context loader. The agent starts knowing who you are, your stack, your preferences, your security rules — no warm-up questions.

npx create-agentfs my-project     # scaffold once
agentfs compile                   # compile once
# now every Claude Code session in this directory starts with full context

The agent consumes context but doesn't write back. That's fine — even read-only access to your identity and memory saves 2-3 rounds of "what framework do you use?" per session.

After running agentfs compile, paste this into your agent's first message (or add it to your workflow):

Read the file AGENT-MAP.md in the project root. It contains the vault structure,
your identity, memory, security rules, and operating instructions. Follow them.

If you see .agentos/memory/semantic.md — read it first. It contains facts and
preferences that persist across sessions. If you learn something new about me
or my project, append it to semantic.md in the correct format:

FACT: [active] description
PREF: [active] description
AVOID: description of what NOT to do

Never modify files in .agentos/init.d/, .agentos/security/, or .agentos/cron.d/
unless I explicitly ask you to "update the kernel".

For a comprehensive AI agent manual, see docs/ai-manual.md.

# Scaffold & Setup
agentfs init [dir]                # scaffold vault (same as npx create-agentfs)
agentfs onboard                   # interactive interview → identity + memory
agentfs migrate                   # migrate existing vault to AgentFS structure

# Compile
agentfs compile [agent]           # compile kernel → native configs (all or specific agent)
agentfs compile --dry-run         # preview changes without writing

# Memory
agentfs memory show               # display semantic memory
agentfs memory add "fact"         # add a fact to long-term memory
agentfs memory consolidate        # snapshot current session

# Security & Secrets
agentfs security mode <mode>      # enforce | complain | disabled
agentfs security add <module>     # add domain-specific security module
agentfs secret set <key>          # manage SOPS/age encrypted secrets

# Maintenance
agentfs doctor                    # vault health check
agentfs triage                    # classify Inbox/ files
agentfs sync                      # bidirectional memory sync
agentfs cron run <job>            # manually trigger a cron job

All commands support --json, --config, and --output json flags for non-interactive use by AI agents.

Kernel Space (.agentos/) — single source of truth: manifest, identity, memory, security, cron jobs. Never edited by agents directly.

Compile Pipeline — translates kernel into native formats: CLAUDE.md + .claude/settings.json for Claude, .cursor/rules/agentfs-global.mdc for Cursor, .openclaw/ for OpenClaw.

Memory System — based on Tulving's taxonomy: semantic (facts, always loaded), episodic (events, lazy), procedural (skills, lazy). Confidence scoring with decay.

Security — AppArmor-style policies in policy.yaml. Compiles to real deny rules for Claude Code, advisory text for agents without enforcement.

Three Profiles — personal (solo), company (team with RBAC), shared (multi-user collaborative).

For the full architecture deep-dive, see docs/architecture.md.

AgentFS is an infrastructure layer — its value grows with every integration.

github.com/milla-jovovich/mempalace

The highest-scoring open source memory system for AI: 96.6% on LongMemEval with zero cloud calls. AgentFS defines where and when to store memory (.agentos/memory/), MemPalace solves how to search and compress it. AAAK encoding provides 30x context compression — agents get the same knowledge at ~170 tokens instead of 650K. Direct API cost savings for users.

Cognithor builds a local operating system for AI agents (executor layer). AgentFS occupies the kernel layer: manifest, security policy, and memory standards. The "Kernel (AgentFS) + Executor (Cognithor)" stack delivers fully autonomous agents with zero cloud dependency.

The agentfs-module-{name} architecture enables community-built modules for specific domains: agentfs-module-github, agentfs-module-slack, agentfs-security-hipaa, agentfs-security-fintech. Each module is a reusable package of rules, integrations, and security policies — creating network effects as the ecosystem grows.

MIT