Give every MCP agent memory across sessions.

ContextDB Cloud exposes its memory tools through the Model Context Protocol. Any MCP-capable agent runtime — coding agents, desktop assistants, orchestration frameworks — gets remember, recall, and the recall_for_action gate without an SDK integration.

Why it matters

Keep customer history when you change models or frameworks.

Teams rarely run one runtime. A support agent in Python, an internal assistant on a desktop client, a workflow bot in TypeScript — each remembers things about the same users. Over MCP they share one memory with one trust policy, and every consequential action passes the same act / ask / abstain gate, recorded in the same decision log.

The same contract

The MCP tools accept exactly what the HTTP routes accept — nothing more. Provenance is required on writes, and the tool descriptions instruct agents not to act on unconfirmed memory.

The same scoping

Every tool call authenticates your project key and runs against a runtime scoped to your org, project, and the user partition in the call. There is no way to name someone else's scope.

An explicit outcome

recall_for_action over MCP returns act, ask, or abstain in the structured result, plus the pending memory IDs an agent should surface for confirmation.

Connect

Connect your MCP client with one configuration.

The endpoint is stateless JSON-RPC over HTTP. Authenticate with your project key; the tools appear in tools/list.

{
  "mcpServers": {
    "contextdb": {
      "url": "https://<your-gateway-host>/mcp",
      "headers": { "Authorization": "Bearer cdb_…" }
    }
  }
}

Tools: remember, recall, recall_for_action, pending_confirmations, confirm. The endpoint handles single JSON-RPC objects and returns JSON; it is part of the hosted alpha and is not production-ready.

The loop

Let the agent ask when a customer detail is unclear.

When the gate answers ask, the agent surfaces the pending fact to the user. On a yes, the agent calls confirm and the fact passes the gate from then on. Confirmation records that an actor attested — it does not prove objective truth, and the decision log keeps the evidence either way.

tools/call recall_for_action {"user_id": "caller-1", "query": "book Friday"}
→ {"outcome": "ask", "memories": [], "pending_confirmation_ids": ["…"]}

# the agent asks; the user says yes

tools/call confirm {"user_id": "caller-1", "memory_id": "…"}
→ {"memory": {"confirmed": true, …}}

tools/call recall_for_action {"user_id": "caller-1", "query": "book Friday"}
→ {"outcome": "act", "memories": […]}

Use the same memory tools in every MCP runtime.

Start with the quickstart, then connect your MCP clients.

Open the quickstart How the gate works