@manasdb/mcp-server

ManasDB = Memory Layer for AI Agents.

AI agents today lack persistent memory. @manasdb/mcp-server provides a plug-and-play memory layer for MCP-compatible assistants like Claude and Cursor — powered by ManasDB's unique polyglot storage architecture.

With a single MCP server, AI agents can:

• 🧠 Store memories across multiple databases simultaneously
• 🔍 Retrieve context with semantic vector search
• 🗑️ Delete outdated knowledge on demand

No custom integration. No boilerplate. Just one command.

---

Quick Start

npx @manasdb/mcp-server setup

This drops you into an interactive setup wizard that generates a ready-to-paste claude_desktop_config.json snippet in under 60 seconds.

---

Architecture

AI Assistant (Claude / Cursor / Any MCP Client)
     │
     │  MCP Protocol (stdio)
     ▼
 @manasdb/mcp-server
     │
     │  Tools: memorize · recall · forget
     ▼
 @manasdb/core  ◄── Embeddings (Transformers / Ollama / Gemini / OpenAI)
     │
 ┌───┴──────────────────┬────────────────────┐
 ▼                      ▼                    ▼
MongoDB             PostgreSQL           Redis / ES
                                         (coming soon)

ManasDB is AI Memory Infrastructure. The components are:

Package	Role
@manasdb/core	Core library: chunking, embedding, polyglot storage
@manasdb/mcp-server	MCP Server: exposes tools to AI assistants
Database Adapters	MongoDB, PostgreSQL (built-in), Redis, Elasticsearch (planned)

---

Why ManasDB?

Feature	ManasDB	Mem0	memGPT	LangChain
Polyglot Storage	✔	✘	✘	✘
Multiple DBs simultaneously	✔	✘	✘	✘
Deduplicated Retrieval	✔	Limited	Limited	Limited
Local Embedding (Ollama)	✔	✘	✘	Limited
Agent Tools (MCP)	✔	Partial	Partial	Partial
Zero Cloud Lock-in	✔	✘	✘	✘

---

Supported Databases

Database	Status	Provider Type
MongoDB (Atlas / Local)	✅ Stable	mongodb
PostgreSQL (pgvector)	✅ Stable	postgres
Redis	🔜 Coming Soon	redis
Elasticsearch	🔜 Coming Soon	elasticsearch

---

Example Interaction

Here is how it works end-to-end in Claude Desktop:

User: Remember my name is Akshay.

Claude → memorize called

{
  "text": "User name is Akshay"
}

✅ Successfully memorized the context across all healthy database providers.

Later in a new conversation...

User: What is my name?

Claude → recall called

{
  "query": "user name"
}

Claude: Your name is Akshay.

---

Setup & Configuration

For deeper information on how ManasDB's polyglot memory, chunking, and retrieval algorithms work, see the primary SDK: @manasdb/core.

Interactive Setup (Recommended)

npx @manasdb/mcp-server setup

The wizard will ask you for:

1. Project Name
2. MongoDB URI
3. PostgreSQL URI
4. Embedding Model Source (transformers, ollama, gemini, openai)
5. Model Name (e.g. nomic-embed-text:latest for Ollama)
6. Enable Telemetry

It then outputs a ready-to-use claude_desktop_config.json snippet.

Manual Configuration

Pass the full config to as a --manas-config argument or set MANAS_DB_CONFIG as an environment variable. The config supports all @manasdb/core options:

{
  "databases": [
    {
      "type": "mongodb",
      "uri": "mongodb://localhost:27017",
      "dbName": "my_app"
    },
    { "type": "postgres", "uri": "postgresql://localhost:5432/postgres" }
  ],
  "projectName": "knowledge_base",
  "modelConfig": { "source": "ollama", "model": "nomic-embed-text:latest" },
  "telemetry": true,
  "debug": false
}

---

Integrations

Claude Desktop

{
  "mcpServers": {
    "manasdb": {
      "command": "npx",
      "args": [
        "-y",
        "@manasdb/mcp-server",
        "--manas-config={\"databases\":[{\"type\":\"mongodb\",\"uri\":\"mongodb://...\"}],\"modelConfig\":{\"source\":\"transformers\"}}"
      ]
    }
  }
}

Cursor

1. Open Settings → Features → MCP Servers
2. Click + Add new MCP Server
3. Type: command
4. Command: npx -y @manasdb/mcp-server --manas-config="{...}"

---

Available Tools

Tool	Description	Input
memorize	Store information into ManasDB	{ "text": "..." }
recall	Retrieve context semantically. Returns results with contentId for each memory.	{ "query": "..." }
forget	Permanently delete a memory from all databases by its contentId (returned by recall).	{ "contentId": "..." }

Tip: Use recall first to find the contentId of the memory you want to delete, then pass it to forget.

---

License

See LICENSE.