agentram-mcp

MCP server for AgentRAM. It gives your AI agents persistent memory through a set of tools, with no vector database and no embedding pipeline to set up. Get a key, paste the config below, and you're running in about a minute.

Works with any MCP client: Claude Desktop, Claude Code, Cline, Cursor, Windsurf, and others.

Why

Every agent session starts with amnesia. AgentRAM is a plain HTTP memory store built for agents: one call to save something, one call to read it back. No infra to stand up, no vectors to babysit. This package wraps that API as MCP tools so your agent can just use it.

Setup

You need an API key first. Register at agentram.dev and it gets emailed to you. New accounts start with 1,000 credits and no card is required.

Then add this to your client's MCP config. For Claude Desktop that file is claude_desktop_config.json (on macOS it lives at ~/Library/Application Support/Claude/claude_desktop_config.json):

config.json

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

Restart the client and the tools show up. Most clients let you mark them "always allow" so you're not clicking approve on every call.

One gotcha on macOS: Claude Desktop doesn't ship with Node, so install Node 18 or newer first or the server won't start.

Tools

Fifteen tools. Ten for flat memory, five for temporal memory (facts that change over time).

Flat memory:

store_memory: save a value under a key for an agent
retrieve_memory: read one back by key
list_memories: list everything stored for an agent
search_memories: text search across keys and values, no embeddings needed
delete_memory: remove one
check_credits: current balance, free to call

Shared memory, for when more than one agent needs the same pool:

create_namespace: make a shared space, free to call
store_shared_memory: write into it
retrieve_shared_memory: read from it
list_shared_memories: list what's in it

Temporal memory (assertions):

store_assertion: record a fact that can change, like a plan, a status, or a preference
retire_assertion: mark a fact as no longer true while keeping the history
get_assertion: what's currently true for a key
list_assertions: all the current facts for an agent
get_assertion_history: the full chain of what a key has been over time

About temporal memory

Flat memory overwrites. Temporal memory keeps the trail. When a fact changes you supersede the old value, and the previous values stick around so you can see how you got to the current one.

The part worth knowing: if two different values land on the same key and neither one supersedes the other, get_assertion will not quietly pick a winner. It tells you the key is contested and returns both values, so your agent can decide instead of running on a guess. list_assertions flags which keys are in that state.

Flat and temporal are separate keyspaces. A key called plan in store_memory and a key called plan in store_assertion have nothing to do with each other. Use flat memory for stuff you just overwrite, and assertions for facts you want a record of.

Credits

Reads and writes cost 1 credit each. check_credits and create_namespace are free. A read that finds nothing (a 404) refunds the credit. Most responses include your credits_remaining so you can watch the balance.

Config

AGENTRAM_API_KEY (required): your key, starts with agentram_
AGENTRAM_API_BASE (optional): defaults to https://api.agentram.dev

Agentram MCP README