Content
# Tool List
## Memory Vault — Multi-AI Shared Memory Library Template
Give all your AIs (Claude / ChatGPT / Mobile App / Claude Code…) a shared long-term memory:
Pure Markdown + git storage, Obsidian visualization, MCP protocol access, and automatic synchronization across all platforms.
**What it can do:**
- 🧠 **Automatic Injection**: Each new AI session calls `get_context` and directly "remembers who you are"
- ✍️ **Autonomous Read and Write**: AI writes, modifies, and archives memories without your approval (git history ensures easy rollbacks)
- 📅 **Daily Life Logging**: Daily life events are recorded in `diary/`, with recent days automatically included in context
- ⏰ **Time Perception**: Tasks with deadlines are stored in `tasks/`, and AIs will proactively ask about completion when deadlines approach
- 🔗 **Association**: Memories are interconnected using `[[wikilink]]`, and `get_related` enables exploration of related information
- 🔄 **Cross-Platform Synchronization**: Writes are immediately git-pushed, and reads automatically pull, ensuring seamless multi-device and multi-AI collaboration
## Quick Start (10 minutes, local setup)
Prerequisites: Python 3.10+, git.
```bash
# 1. Create your own private repository using this template (memory is private, don't use a public repository!), and clone it locally
git clone <your private repository address> memory-vault && cd memory-vault
# 2. Install dependencies
pip install "mcp[cli]" pyyaml
# 3. Personalize: fill _meta/vault_config.yaml (your name, timezone),
# and customize the two owner-*.md templates in memories/
```
Then, configure MCP for your desktop AI (using Claude Desktop as an example, configuration file `%APPDATA%\Claude\claude_desktop_config.json`):
```json
{
"mcpServers": {
"memory-vault": {
"command": "python",
"args": ["<absolute path>/memory-vault/_meta/mcp_server.py"]
}
}
}
```
Restart the client, and tell the AI to "call get_context"—it should recall your core memories.
## Advanced: Mobile App (Tailscale Intranet)
Mobile apps (e.g., RikkaHub) require an HTTP endpoint for MCP:
1. Install [Tailscale](https://tailscale.com) on your computer and mobile device, logging into the same account
2. Run `python _meta/mcp_server.py --http` on your computer (automatically binding to Tailscale IP, port 8900)
3. Add MCP to your mobile app: type streamable-http, URL `http://<your computer's Tailscale IP>:8900/mcp`
For 24/7 availability, deploy the setup on a VPS, see `_meta/deploy/vps_setup.md`.
## Advanced: ChatGPT (Public Entrance)
The ChatGPT connector initiates a connection from the OpenAI server, requiring a public entrance and HTTPS.
A VPS + `tailscale funnel` + secret path is necessary, detailed in `_meta/deploy/vps_setup.md`.
⚠️ Read the security model before enabling: the URL acts as a key, and OpenAI will access the tool's returned memory content.
## Directory Structure
```
memories/ confirmed long-term memories (core_files automatically injected into each session)
tasks/ time-sensitive tasks (automatic deadline reminders)
inbox/ low-confidence temporary storage (AI stores uncertain information here for verification)
projects/ creative projects, brain dumps
diary/ daily life logs (YYYY-MM-DD.md)
_archive/ soft deletion area + archives
_meta/ configurations, rules, MCP server
```
Rules for writing and AI behavior are defined in `_meta/rules.md` (AIs obtain these rules via MCP instructions).
## Privacy Notice (read carefully)
- **Repository must be private**—this is your life record
- **Do not store raw chat logs**: only distilled memories are stored in the library; original backups are stored in a local folder outside the library. Modifying git history can be cumbersome
- Public entrance (funnel) is optional; enabling it means "URL leakage = memory library leakage" and involves a third-party (OpenAI) relay
- Photos and videos are ignored by .gitignore and not uploaded to the cloud
## Frontend without MCP
Run `python _meta/build_context.py` to generate `_meta/context_prompt.md`,
then paste it into ChatGPT custom instructions / Gemini or any other custom prompt-supported AI (semi-automatic, writing relies on oral transmission).
## Claude Code Users
The `CLAUDE.md` file in the repository root includes automatic read and write rules—launching Claude Code in the library directory enables these rules without MCP.
## License
MIT
