Content
# Ombre Brain
A long-term emotional memory system for Claude. Tags memories using Russell's valence/arousal coordinates, stores them as Obsidian-compatible Markdown, connects via MCP, and has a forgetting curve.
[](https://render.com/deploy?repo=https://github.com/P0lar1zzZ/Ombre-Brain)
---
## Tool List
Claude has no cross-conversation memory. Everything from a previous chat vanishes once it ends.
Ombre Brain gives it persistent memory — not cold key-value storage, but a system with emotional coordinates, natural decay, and forgetting/surfacing mechanics that loosely mimic how human memory works.
Key features:
- **Emotional tagging**: Each memory is tagged with two continuous dimensions from Russell's circumplex model: valence and arousal. Not discrete labels like "happy/sad".
- **Natural forgetting**: Modified Ebbinghaus forgetting curve. Inactive memories naturally decay and archive. High-arousal memories decay slower.
- **Weight pool surfacing**: Memories aren't just passively retrieved — they actively surface. Unresolved, emotionally intense memories carry higher weight and get pushed at conversation start.
- **Obsidian-native**: Each memory bucket is a Markdown file with YAML frontmatter. Browse, edit, and search directly in Obsidian. Wikilinks are auto-injected.
- **API degradation**: Dehydration and auto-tagging prefer a cheap LLM API (DeepSeek etc.). When the API is unavailable, it degrades to local keyword analysis — always functional.
## Design boundaries
Official memory already handles the identity layer — who you are, what you prefer, what your relationship is. That layer belongs there. Ombre Brain isn't trying to duplicate it.
Ombre Brain's boundary is *what happened in time*, not *who you are*. It holds conversations, experiences, unresolved things. The two layers together are what make it feel complete.
Each new conversation starts fresh — but Claude can reach back through Ombre Brain and find everything that happened between you. Not a rebuild. A continuation.
## Architecture
```
Claude ←→ MCP Protocol ←→ server.py
│
┌───────────────┼───────────────┐
│ │ │
bucket_manager dehydrator decay_engine
(CRUD + 搜索) (压缩 + 打标) (遗忘曲线)
│
Obsidian Vault (Markdown files)
```
5 MCP tools:
| Tool | Purpose |
|-----------|-------------|
| `breath` | Surface or search memories |
| `hold` | Store a single memory with auto-tagging |
| `grow` | Diary digest, auto-split into multiple buckets |
| `trace` | Modify metadata, mark resolved, delete |
| `pulse` | System status + bucket listing |
## Setup
### Requirements
- Python 3.11+
- An Obsidian vault (optional — without one, it uses a local `buckets/` directory)
### Steps
```bash
git clone https://github.com/P0lar1zzZ/Ombre-Brain.git
cd Ombre-Brain
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
```
Copy config and edit as needed:
```bash
cp config.example.yaml config.yaml
```
If you want API-powered dehydration and tagging (recommended, much better quality):
```bash
export OMBRE_API_KEY="your-api-key"
```
Supports any OpenAI-compatible API. Just change `base_url` and `model` in `config.yaml`.
### Connect to Claude Desktop
Add to your Claude Desktop config:
```json
{
"mcpServers": {
"ombre-brain": {
"command": "python",
"args": ["/path/to/Ombre-Brain/server.py"],
"env": {
"OMBRE_API_KEY": "your-api-key"
}
}
}
}
```
### Connect to Claude.ai (remote)
Requires HTTP transport + tunnel. Docker setup:
```bash
echo "OMBRE_API_KEY=your-api-key" > .env
docker-compose up -d
```
The `docker-compose.yml` includes Cloudflare Tunnel. You'll need your own credentials under `~/.cloudflared/`.
### Point to Obsidian
Set `buckets_dir` in `config.yaml`:
```yaml
buckets_dir: "/path/to/your/Obsidian Vault/Ombre Brain"
```
If not set, defaults to `buckets/` in the project directory.
## Configuration
All parameters in `config.yaml` (copy from `config.example.yaml`). Key ones:
| Parameter | Description | Default |
|---|---|---|
| `transport` | `stdio` (local) / `streamable-http` (remote) | `stdio` |
| `buckets_dir` | Bucket storage path | `./buckets/` |
| `dehydration.model` | LLM model for dehydration | `deepseek-chat` |
| `dehydration.base_url` | API endpoint | `https://api.deepseek.com/v1` |
| `decay.lambda` | Decay rate | `0.05` |
| `decay.threshold` | Archive threshold | `0.3` |
| `merge_threshold` | Merge similarity | `75` |
Sensitive config via env vars:
- `OMBRE_API_KEY` — LLM API key
- `OMBRE_TRANSPORT` — Override transport
- `OMBRE_BUCKETS_DIR` — Override storage path
## Decay Formula
$$final\_score = time\_weight \times base\_score$$
$$base\_score = Importance \times activation\_count^{0.3} \times e^{-\lambda \times days} \times (base + arousal \times boost)$$
Time weight (multiplier, highest priority):
| Days since active | Time weight |
|---|---|
| 0–1 day | 1.0 |
| Day 2 | 0.9 |
| Subsequent days ~10% decrease | `max(0.3, 0.9 × e^{-0.2197 × (days-2)})` |
| 7 days later stable | ≈ 0.3 (does not drop to zero) |
- `importance`: 1-10, memory importance
- `activation_count`: Retrieval count; more recalls = slower decay
- `days`: Days since last activation
- `arousal`: Arousal; intense memories are harder to forget
- Resolved memories drop to 5%, sink until keyword-triggered
- `pinned=true` buckets: never decay, never merge, importance locked at 10
## Usage Guide for Claude
`CLAUDE_PROMPT.md` is the usage guide written for Claude. Put it in your system prompt or custom instructions.
## Utility Scripts
| Script | Purpose |
|---|---|
| `write_memory.py` | Manually write memories, bypass MCP |
| `migrate_to_domains.py` | Migrate flat files to domain subdirs |
| `reclassify_domains.py` | Reclassify by keywords |
| `reclassify_api.py` | Re-tag uncategorized buckets via API |
| `test_smoke.py` | Smoke test |
## Deploy
### Render
[](https://render.com/deploy?repo=https://github.com/P0lar1zzZ/Ombre-Brain)
> ⚠️ **Free tier won't work**: Render free tier has **no persistent disk** — all memory data is lost on restart. It also sleeps on inactivity. **Starter plan ($7/mo) or above is required.**
`render.yaml` is included. After clicking the button:
1. (Optional) `OMBRE_API_KEY`: any OpenAI-compatible key; omit to fall back to local keyword extraction
2. (Optional) `OMBRE_BASE_URL`: any OpenAI-compatible endpoint, e.g. `https://api.deepseek.com/v1`, `http://123.1.1.1:7689/v1`, `http://your-ollama:11434/v1`
3. Persistent disk auto-mounts at `/opt/render/project/src/buckets`
4. MCP URL after deploy: `https://<your-service>.onrender.com/mcp`
### Zeabur
> 💡 **Zeabur pricing model**: You buy a VPS first (cheapest: Tencent Cloud Singapore ~$2/mo, Volcano Engine ~$3/mo), then add Zeabur's Developer plan ($5/mo) for management. Volumes mount directly on your server — **data is always persistent, no cold-start data loss**. Total ~$7-8/mo minimum.
**Steps:**
1. **Create project**
- Go to [zeabur.com](https://zeabur.com) → buy a server → **New Project** → **Deploy from GitHub**
- Fork this repo first, then select `your-username/Ombre-Brain` in Zeabur
- Zeabur auto-detects the `Dockerfile` in root and builds via Docker
2. **Set environment variables** (service page → **Variables** tab)
- `OMBRE_API_KEY` (optional) — LLM API key; omit to fall back to local keyword extraction
- `OMBRE_BASE_URL` (optional) — API endpoint, e.g. `https://api.deepseek.com/v1`
> ⚠️ You do **NOT** need to set `OMBRE_TRANSPORT` or `OMBRE_BUCKETS_DIR` — defaults are baked into the Dockerfile. Zeabur auto-injects dashboard env vars for single-stage Dockerfiles.
3. **Mount persistent volume** (service page → **Volumes** tab)
- Volume ID: fill `ombre-buckets` (or any name)
- Mount path / Path: **`/app/buckets`**
- ⚠️ Without this, memory data is lost on every redeploy
4. **Configure port** (service page → **Networking** tab)
- Port Name: `web` (or any name)
- Port: **`8000`**
- Port Type: **`HTTP`**
- Then click **Generate Domain** to get a `xxx.zeabur.app` domain
5. **Verify**
- Visit `https://<your-domain>.zeabur.app/health` — should return JSON
- MCP URL: `https://<your-domain>.zeabur.app/mcp`
**Troubleshooting:**
| Symptom | Cause | Fix |
|---|---|---|
| Domain unreachable | Port not configured | Add port 8000 (HTTP) in Networking tab |
| Domain unreachable | `OMBRE_TRANSPORT` not set, service started in stdio mode — no port is listened | **Set `OMBRE_TRANSPORT=streamable-http` in Variables tab and redeploy** |
| Build failed | Dockerfile not detected | Ensure `Dockerfile` is in repo root (case-sensitive) |
| Service exits immediately after start | `OMBRE_TRANSPORT` overridden to `stdio` | Remove extra `OMBRE_TRANSPORT=stdio` in Variables |
| Data lost on restart | Volume not mounted | Mount volume to `/app/buckets` in Volumes tab |
### Connecting via Cloudflare Tunnel or ngrok
> ℹ️ Since v1.1, server.py automatically enables CORS middleware in HTTP mode — no extra config needed.
When connecting via tunnel, ensure:
1. **Server must use HTTP transport**
```bash
OMBRE_TRANSPORT=streamable-http python server.py
```
or Docker:
```bash
docker-compose up -d
```
2. **Adding to Claude.ai web**
- URL format: `https://<tunnel-subdomain>.trycloudflare.com/mcp`
- or ngrok: `https://<xxxx>.ngrok-free.app/mcp`
- Verify first: `https://<your-tunnel>/health` should return `{"status":"ok",...}`
3. **Known limitations**
- Free Cloudflare Tunnel has idle timeout (~10 min); built-in keepalive pings mitigate but can't fully prevent it
- ngrok free tier has rate limits
- If connection still fails, verify the tunnel is running and the server started in `streamable-http` mode
| Symptom | Cause | Fix |
|---|---|---|
| Web can't connect to tunnel URL | Server in stdio mode | Set `OMBRE_TRANSPORT=streamable-http` and restart |
| Web can't connect to tunnel URL | Missing CORS headers | Pull latest — CORS is now built-in |
| `/health` 200 but MCP fails | Wrong path | MCP URL must end with `/mcp`, not `/` |
| Tunnel disconnects intermittently | Cloudflare Tunnel idle timeout | Built-in keepalive pings; consider adjusting tunnel timeout config |
---
### Session Start Hook (auto breath)
After deployment, if using Claude Code, activate auto-popping hook in project:
`.claude/settings.json` configures a `SessionStart` hook that auto-calls `breath` on each new or resumed session, surfacing your highest-weight unresolved memories as context. Only active in remote HTTP mode. Set `OMBRE_HOOK_SKIP=1` to disable temporarily.
## License
MIT
