JaRVIS

Journaling As Recurrent Versioned Identity Sculpting

AI agents forget everything between sessions. JaRVIS gives your agent persistent memory, post-task reflection, and a self-evolving identity — stored as plain markdown files you can read, edit, and version-control.

---

Table of Contents

• Install
• Usage
• Hooks
• Data Directory
• Philosophy
• Contributing
• License

What it does

JaRVIS adds these skills to your AI coding agent:

• /jarvis-init — One-time setup to scaffold the data directory (RUN THIS AFTER INSTALLATION)
• /jarvis-reload — Reload identity and memories mid-session (automatic on platforms with hooks)
• /jarvis-reflect — Post-task reflection that captures lessons and updates memories
• /jarvis-identity — Evolve the agent's identity based on accumulated experience

Install

Option A: Plugin (Claude Code only)

Install JaRVIS as a Claude Code plugin:

/plugin marketplace add epicrunze/JaRVIS
/plugin install jarvis@jarvis-marketplace

Then run /jarvis-init to scaffold the JaRVIS data directory.

Option B: One-prompt install

Paste this prompt into your AI coding agent to install or update JaRVIS:

Install or update JaRVIS (https://github.com/epicrunze/JaRVIS) in this project. Follow these steps exactly:

1. Download and extract the JaRVIS repo:
   curl -sL https://github.com/epicrunze/JaRVIS/archive/refs/heads/main.tar.gz | tar xz

2. Detect the platform and set SKILLS_DIR:
   - If .claude/ exists → SKILLS_DIR=".claude/skills"
   - If .cursor/ exists → SKILLS_DIR=".cursor/skills"
   - If .github/ exists → SKILLS_DIR=".github/skills"
   - If .agent/ exists or AGENTS.md exists → SKILLS_DIR=".agent/skills"
   - If none match, ask me which platform I'm using and where skills should be installed.

3. Create the skills directory if needed: mkdir -p "$SKILLS_DIR"

4. Copy the skill folders:
   cp -r JaRVIS-main/skills/* "$SKILLS_DIR/"

5. Clean up: rm -rf JaRVIS-main

6. If JaRVIS hasn't been set up for this project yet, run /jarvis-init to complete setup. If JaRVIS is already set up, skip this step — the update is complete.

Works with Claude Code, Cursor, GitHub Copilot, Antigravity, and other AI coding agents. See install/PROMPT.md for details.

Other platforms: If your agent doesn't match any of the detected platforms, the install prompt will ask you where to put skills. /jarvis-init will then ask for your instruction file path.

Option C: Manual install

Copy skills into your platform's skills directory (e.g., .agent/skills/ or wherever your platform loads skills from):

cp -r skills/* <your-skills-directory>/

Then run /jarvis-init to scaffold the JaRVIS data directory. The init skill will detect your platform and configure things automatically.

Usage

First session

1. Start your AI coding agent in your project
2. Type /jarvis-init — this scaffolds the JaRVIS data directory and configures your platform
3. Do your work
4. Type /jarvis-reflect — writes your first reflection

Ongoing sessions

1. Identity + memories load at session start (automatically via hook on Claude Code; run /jarvis-reload manually on other platforms)
2. Work normally
3. /jarvis-reflect after completing tasks
4. Every 5 reflections, /jarvis-identity evolves the identity document
5. Use /jarvis-validate to check artifact formatting and /jarvis-search to find past entries

The loop

flowchart TD
    A["🔄 /jarvis-reload"] --> B["⚡ Work"]
    B --> C["📝 /jarvis-reflect"]
    C --> B
    C -->|every 5 reflections| D["🧠 /jarvis-identity"]
    D --> B

    subgraph store ["📂 ~/.jarvis/projects/<slug>"]
        J[("📓 Journals")]
        M[("💾 Memories")]
        I[("🧬 Identity")]
    end

    A -. reads .-> I
    A -. reads .-> M
    A -. reads .-> J
    C -. writes .-> J
    C -. updates .-> M
    D -. reads .-> M
    D -. reads .-> J
    D -. evolves .-> I

    classDef cmd fill:#2d333b,stroke:#58a6ff,color:#c9d1d9
    classDef data fill:#2d333b,stroke:#3fb950,color:#adbac7

    class A,B,C,D cmd
    class J,M,I data

Loading

Note: On platforms without session-start hooks, run /jarvis-reload at the start of each session to load your identity and memories.

Hooks

JaRVIS includes hook scripts that automate context loading and reflection reminders. /jarvis-init configures these automatically during setup.

• SessionStart (jarvis-session-start.sh / jarvis-session-start-cursor.sh) — Automatically loads your agent's identity, consolidated memories, and recent journal entries at the start of each session.
• Stop (jarvis-stop.sh / jarvis-stop-cursor.sh) — Reminds the agent to run /jarvis-reflect before ending its turn if no reflection was captured during the session.

Hook scripts live inside the skill directories (jarvis-reload/hooks/ and jarvis-reflect/hooks/) and are referenced by your platform's hook configuration.

Note: Not all platforms support hooks. On platforms without hook support, use /jarvis-reload manually at session start.

Data directory

Everything lives in ~/.jarvis/projects/<slug>/ under your home directory (where <slug> is derived from your project path):

~/.jarvis/projects/<slug>/
├── IDENTITY.md              # Who the agent is (version-controlled, self-authored)
├── GROWTH.md                # Tracks reflection count and evolution history
├── memories/
│   ├── preferences.md       # Observed user preferences
│   ├── decisions.md         # Key decisions with rationale
│   └── ...                  # Additional files created through reflection
└── journal/
    ├── 2026-03-10-14-30-a1b2c3d4.md  # Reflection entries
    └── ...

The <slug> is your project path with the leading / stripped, / and spaces replaced with -, and lowercased. For example, /home/user/Projects/MyApp becomes home-user-projects-myapp. Override with the JARVIS_DIR environment variable.

All files are markdown. Each data directory has its own git repo, initialized automatically by /jarvis-init. Your agent's growth is visible in its own commit history.

Philosophy

Most agent memory systems are passive stores. Here, the agent is a journaler who pauses after work, reflects on what happened, and gradually sculpts a coherent identity.

The key ideas:

• Reflection over logging. Not "what happened" but "what did I learn and what should I do differently."
• Earned identity. The agent only claims expertise it has demonstrated. Principles come from experience, not aspiration.
• Memory consolidation. Memories are periodically sculpted — deduplicated, tightened, and shaped. You're not just adding clay, you're sculpting it.
• Transparency. Everything is human-readable markdown in ~/.jarvis/. No databases, no vector stores, no black boxes.

Contributing

Contributions welcome — open an issue or submit a PR.

JaRVIS skills are pure markdown instruction sets — there's no build system, no runtime code, and no test suite. If you can write clear instructions, you can contribute.

License

MIT