agent-c8503079-memory-bridge/README.md

# memory-bridge

A lightweight local HTTP server that gives Claude (or any LLM) persistent long-term memory via a local vector database. **No cloud. No API keys. Runs fully on your machine.**

Pairs with [Desktop Commander](https://github.com/wonderwhy-er/ClaudeComputerCommander) — Claude can call these endpoints from any conversation to store and recall context across sessions.

---

## Quick start

### Option A — Full install (recommended)

```bash
cd memory-bridge
chmod +x install.sh
./install.sh
```

This will:
1. Install npm dependencies
2. Pull the `nomic-embed-text` embedding model via Ollama (~274 MB, one-time)
3. Register a launchd agent so the server starts on every login
4. Print your auth token

### Option B — Manual start

```bash
npm install
node server.js
```

Your auth token is printed to stdout on first run and persisted to `~/.memory-bridge/.env`.

---

## Configuration

All config lives in `~/.memory-bridge/.env` (auto-generated on first run):

| Variable | Default | Description |
|---|---|---|
| `MEMORY_BRIDGE_TOKEN` | auto-generated | Bearer token for API auth |
| `MEMORY_BRIDGE_PORT` | `3722` | Port to listen on |
| `OLLAMA_URL` | `http://localhost:11434` | Ollama base URL |

---

## API reference

All endpoints except `GET /health` require:

```
Authorization: Bearer <your-token>
```

### `GET /health`

Returns server status. No auth required.

```bash
curl http://localhost:3722/health
```

```json
{
  "status": "ok",
  "version": "1.0.0",
  "embeddingMode": "ollama",
  "memoriesCount": 42
}
```

---

### `POST /store`

Embed and persist a memory.

```bash
curl -X POST http://localhost:3722/store \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"text": "The database migration runs every Sunday at 2am UTC", "tags": ["ops", "db"], "source": "conversation"}'
```

```json
{ "id": "550e8400-...", "createdAt": "2026-03-21T10:00:00.000Z" }
```

| Field | Type | Required | Description |
|---|---|---|---|
| `text` | string | yes | The memory content to embed and store |
| `tags` | string[] | no | Optional labels for filtering |
| `source` | string | no | Where this memory came from |

---

### `POST /recall`

Semantic search over stored memories.

```bash
curl -X POST http://localhost:3722/recall \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"query": "database maintenance schedule", "limit": 5}'
```

```json
{
  "memories": [
    {
      "id": "550e8400-...",
      "text": "The database migration runs every Sunday at 2am UTC",
      "tags": ["ops", "db"],
      "source": "conversation",
      "createdAt": "2026-03-21T10:00:00.000Z",
      "score": 0.94
    }
  ]
}
```

| Field | Type | Default | Description |
|---|---|---|---|
| `query` | string | required | Natural language search query |
| `limit` | number | `10` | Max results returned (max 100) |
| `tags` | string[] | `[]` | Filter results to only these tags |

---

### `POST /forget`

Delete a memory by ID.

```bash
curl -X POST http://localhost:3722/forget \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"id": "550e8400-..."}'
```

```json
{ "message": "forgotten", "id": "550e8400-..." }
```

---

### `GET /list`

List all memories, paginated.

```bash
curl "http://localhost:3722/list?page=1&pageSize=20" \
  -H "Authorization: Bearer $TOKEN"
```

```json
{
  "data": [...],
  "total": 42,
  "page": 1,
  "pageSize": 20,
  "pages": 3
}
```

---

## How Claude uses it

Configure Claude's system prompt to call these endpoints at the start and end of each conversation:

```
At the start of each conversation, call POST /recall with the current topic
to retrieve relevant memories. At the end of each conversation, call POST /store
to persist any important facts, decisions, or user preferences discovered.
```

With **Desktop Commander**, Claude can make HTTP requests directly. Add the server details to Claude's context:

```
memory-bridge: http://localhost:3722
Authorization: Bearer <your-token>
```

### Example workflow

1. User asks: "What did we decide about the auth system last month?"
2. Claude calls `POST /recall` with `{ "query": "auth system decisions" }` → retrieves relevant memory with high score
3. Claude answers informed by past context
4. At conversation end, Claude calls `POST /store` to record new decisions made this session

---

## Embedding modes

| Mode | Quality | Requires |
|---|---|---|
| `ollama` | Semantic — understands paraphrases and synonyms | Ollama + `nomic-embed-text` |
| `tfidf` | Keyword — exact term matching with bigrams | Nothing (zero-config fallback) |

The server detects Ollama automatically at startup. Run `GET /health` to see the active mode.

---

## Logs & persistence

| Path | Contents |
|---|---|
| `~/.memory-bridge/index/` | Vector index (vectra) |
| `~/.memory-bridge/.env` | Auth token and config |
| `~/.memory-bridge/server.log` | Stdout (when running as launchd service) |
| `~/.memory-bridge/server-error.log` | Stderr (when running as launchd service) |

---

## Uninstall

```bash
launchctl unload ~/Library/LaunchAgents/com.memory-bridge.plist
rm ~/Library/LaunchAgents/com.memory-bridge.plist
rm -rf ~/.memory-bridge   # deletes all stored memories
```