amber

# Amber

**Long-term memory for AI assistants.**

Amber is an MCP server that gives any AI assistant persistent, searchable memory across conversations. Your AI remembers preferences, decisions, project context, and personal details - without you doing anything special.

> Just talk normally. Amber stores what matters and finds it when relevant.

## Quick Install

One command. Works with any MCP-compatible client.

### Claude Code / Claude Desktop

```bash
claude mcp add --transport http --scope user amber https://mcp.ambermem.com
```

### Cursor

Add to `~/.cursor/mcp.json` (or `%USERPROFILE%\.cursor\mcp.json` on Windows):

```json
{
  "mcpServers": {
    "amber": {
      "url": "https://mcp.ambermem.com"
    }
  }
}
```

### ChatGPT

Settings → Connectors → Create → URL: `https://mcp.ambermem.com`

### Windsurf

Add to `~/.codeium/windsurf/mcp_config.json`:

```json
{
  "mcpServers": {
    "amber": {
      "serverUrl": "https://mcp.ambermem.com"
    }
  }
}
```

### VS Code (GitHub Copilot)

Add to `.vscode/mcp.json`:

```json
{
  "servers": {
    "amber": {
      "type": "http",
      "url": "https://mcp.ambermem.com"
    }
  }
}
```

### Any MCP client

URL: `https://mcp.ambermem.com` | Transport: Streamable HTTP | Auth: OAuth 2.1 (auto-discovered)

## How It Works

1. **You talk to your AI normally.** Amber stores important facts in the background.
2. **Next conversation,** your AI searches Amber automatically when context would help.
3. **Memory improves over time.** The more you use it, the better it gets.

No configuration. No tagging. No manual organization.

## What Makes Amber Different

| Feature | Basic memory servers | Amber |
|---------|---------------------|-------|
| Storage | One embedding per memory | **Multiple semantic variants** per fact |
| Search | Single vector lookup | **Hybrid: vector + keyword + RRF fusion** |
| Queries | Exact match only | **Many phrasings per fact**, matched semantically |
| Input | Stored as-is | **Chunked into atomic facts**, each independently searchable |
| Topics | Manual tags or none | **Auto-grouped**, matched semantically at search time |
| Privacy | Varies by server | **No generative model reads your memories** |
| Time | No temporal awareness | **Natural language time parsing** ("last week", "3 days ago") |

## Technical Details

- **29 MCP tools** (14 memory, 13 account, 2 feedback/notification)
- **Hybrid retrieval pipeline**: vector search + full-text search + Reciprocal Rank Fusion
- **Multi-variant embeddings**: every fact is stored with several paraphrases — at least 5, and a store is rejected below that — which is the main thing that makes it findable later
- **No generative model reads your memories**: your assistant does the chunking, the topics and the phrasings. Memory text is sent only to an *embedding* endpoint (a vector model, not a chat model), so no chat model is ever shown what you store
- **Atomic facts**: a conversation becomes individually searchable facts rather than one blob, each with its own subjects, topics and dates
- **Temporal search**: "what did I say last week?" resolves to a real date range rather than a keyword match
- **Automatic topic grouping**: memories are grouped by topic, and a search for "work" also finds "career" and "job"
- **Async processing**: storage completes in the background, never blocking your conversation

## Pricing

- **90-day free trial** - no charge, cancel anytime
- **$2.99/month** after trial, via PayPal
- **Cancel instantly** - ask your AI to cancel, or cancel through PayPal directly
- **No lock-in** - export all your data as JSON anytime

## Privacy

- No email collected
- No marketing, no spam
- Data isolated per user (separate database)
- PayPal handles all payment info
- Full export + account deletion available
- GDPR compliant (data minimization by design)

## Architecture

Amber runs on Cloudflare Workers (zero cold starts, global edge deployment) with Turso databases (one per user, full isolation). Your assistant does the chunking, the topics and the phrasings, so no generative model is ever shown what you store; memory text is sent only to OpenAI's embedding endpoint.

For full technical documentation: [ambermem.com/llms.txt](https://ambermem.com/llms.txt)

## Links

- **Website**: [ambermem.com](https://ambermem.com)
- **MCP endpoint**: `https://mcp.ambermem.com`
- **Privacy policy**: [ambermem.com/privacy](https://ambermem.com/privacy)
- **Terms of service**: [ambermem.com/terms](https://ambermem.com/terms)
- **Technical docs (for AI)**: [ambermem.com/llms.txt](https://ambermem.com/llms.txt)
- **Report a problem**: [ambermem.com/report](https://ambermem.com/report) — no account needed, and it works even when your client cannot connect. In-session, an assistant can also call the `amber_send_feedback_to_developer` tool directly.

## FAQ

**Will it slow my AI down?**
No. Storage is async (background). Search adds <1 second.

**What if Amber shuts down?**
Export all your data as JSON anytime. Your data is always yours.

**Do I need a PayPal account?**
Currently yes. PayPal handles both identity and billing. More login options coming soon.

**Is my data safe?**
Each user gets a completely isolated database. No data is shared between users. Amber has no access to your PayPal payment details.

**Can I self-host?**
Not currently. Amber is a managed service. We handle the infrastructure, scaling, and LLM costs so you don't have to.