AgentRAM + MCP

AgentRAM ships an MCP server, so any MCP-compatible client, Claude Code, Codex, Cursor, Cline, Claude Desktop, the OpenAI Agents SDK, can give its agent persistent memory with a single config block. This is the shared foundation the other integration guides build on.

What you get

• Persistent memory across sessions. Values you store stay available in every future session, not just the current context window.
• Deterministic recall by key. You ask for a named value and get exactly that value back. No similarity thresholds, no guessing.
• Shared memory for agent teams. Several agents can read and write the same namespace to stay in sync.
• No vector database. No embeddings, no index to provision, nothing extra to run or pay for.
• Memory as native tools. Your agent calls store and recall as first-class tools, the same way it calls anything else.

How it works

MCP, the Model Context Protocol, is a standard way for agent clients to load external tools. AgentRAM publishes an MCP server, agentram-mcp, that exposes its memory operations, store, recall, list, search, delete, and shared memory, as tools. You register the server once, and from then on your agent can remember and recall as part of its normal tool use.

Because it is standard MCP, the same server works in Claude Code, Codex, Cursor, Cline, Claude Desktop, and the OpenAI Agents SDK. The only thing that changes between them is where the config lives, which the client guides below cover.

Prerequisites

• An AgentRAM API key. Create one free: 1,000 operations, no card.
• Node.js 18 or newer, so the server can run through npx. Check with node -v.

Configure the server

Every MCP client needs the same three things: the command, its arguments, and an environment variable holding your API key. The generic entry looks like this:

{
  "mcpServers": {
    "agentram": {
      "command": "npx",
      "args": ["-y", "agentram-mcp"],
      "env": { "AGENTRAM_API_KEY": "agentram_your_key_here" }
    }
  }
}

Keep your key in an environment variable rather than committing it. Where this block lives depends on your client.

Add it to your client

• Claude Code
• Codex
• OpenAI Agents SDK

Using Cursor, Cline, or Claude Desktop? They read the same mcpServers block shown above in their MCP settings.

Quick start

Once the server is connected, your agent has memory tools. Ask it to store something:

Store my preferred language as "French" under the key user_language.

Then, in a brand new session, ask it to recall:

What is my preferred language? Recall user_language from AgentRAM.

It reads the value back from AgentRAM instead of starting from nothing. That is the whole point: memory that outlives the session.

Smoke test and troubleshooting

To confirm the connection, ask your agent to list the AgentRAM tools it can see, then store and recall one test value. If it reads back the same value, you are done.

• Tools do not appear. Reload or restart the client after adding the config. Confirm the block is valid and in the right file for your client.
• 401 or auth error. The key is missing or wrong. Check AGENTRAM_API_KEY is set and starts with agentram_.
• npx cannot find agentram-mcp. Node is older than 18, or you are offline. Check node -v.