Skip to content

atomsai/contextdb-memory

v1.1.0Apache-2.0

Customer memory with recall, confirmation, forgetting, and pre-action evidence checks.

GitHubOpen Website

MCP Customer Memory Connector for AI Agents — ContextDB

ContextDB is an OAuth MCP connector that gives ChatGPT, Claude, Cursor, Grok, Lovable, and other MCP clients six scoped customer-memory tools. It remembers sourced facts, recalls context, checks evidence before actions, records confirmation, and forgets selected memory without exposing a project-wide API key to the client.

The old workflow copied a server credential into every AI client and let the caller choose arbitrary customer partition IDs. ContextDB OAuth grants instead bind one account, project, scope set, and opaque memory partition on the server.

Install in Cursor

Add ContextDB to Cursor

Or add the remote server manually:

{
  "mcpServers": {
    "contextdb": {
      "url": "https://api.contextdb.ai/mcp"
    }
  }
}

Cursor opens ContextDB authorization on first connection. Sign in, choose one project, review the scopes, and approve. No cdb_ project key belongs in this file.

Other clients

Claude-compatible clients can load the repository's .mcp.json.

Grok Build can load .grok/config.toml or add the server directly:

grok mcp add --transport http contextdb https://api.contextdb.ai/mcp

For ChatGPT, Claude.ai, Grok web, or Lovable, add this remote connector URL:

https://api.contextdb.ai/mcp

Each client should discover ContextDB's OAuth metadata and open the same project-consent flow.

Try the tools

After authorization, ask your AI client:

Remember that I prefer email updates.
How should you contact me?
Check whether confirmed memory supports booking Friday.
Show memory that still needs confirmation.
Forget the email preference you just stored.

The connector exposes:

  • remember
  • recall
  • recall_for_action
  • pending_confirmations
  • confirm
  • forget

OAuth tool schemas do not accept user_id. ContextDB derives a stable opaque partition from the authorizing account and project.

Use cases

  • Voice-agent developers carry a caller's confirmed preferences across calls.
  • Support engineering leads ground replies in sourced customer context.
  • Safety and governance teams call recall_for_action before bookings, refunds, or account changes and preserve act, ask, or abstain.

ContextDB advises. The customer application remains responsible for business authorization, current system state, and final action execution.

Skill: compact your context safely

Agents that compact, summarize, or rewrite their own context during long tasks can lose what the user said or start acting on their own summary. The contextdb-compaction skill tells the agent to save customer facts with honest source labels before it compacts, to store facts rather than instructions, and to call recall_for_action before acting afterwards.

Install it for Cursor or Claude Code:

mkdir -p ~/.cursor/skills/contextdb-compaction
curl -fsSL https://raw.githubusercontent.com/atomsai/contextdb-mcp-plugin/main/skills/contextdb-compaction/SKILL.md \
  -o ~/.cursor/skills/contextdb-compaction/SKILL.md

For Claude Code, use ~/.claude/skills/contextdb-compaction/ instead. Any client that reads Agent Skills SKILL.md files can load the same file.

The compaction proof runs offline on the Apache-2.0 SDK and shows what ContextDB does with the notes the skill saves:

  • the customer's own words get act;
  • the agent's own summary gets ask;
  • an instruction written into the agent's notes gets abstain.

We have not yet measured how consistently agents follow the skill. Treat it as guidance and keep the action check in your application.

Security and compatibility

  • OAuth 2.1 authorization code with S256 PKCE
  • One-hour access tokens
  • Rotating 30-day refresh tokens
  • Immediate grant and token revocation
  • Exact project consent
  • Read/write scope filtering
  • Destructive-tool annotations
  • Server-bound memory partition

The server uses stateless Streamable HTTP with JSON responses and MCP protocol 2025-06-18. It is tools-only: no SSE, sessions, resumability, resources, prompts, or server-initiated messages.

Current status

ContextDB Cloud and the OAuth MCP connector are Hosted Alpha with no availability SLA. The server is published in the official MCP Registry as io.github.atomsai/contextdb-memory@0.1.0. The ChatGPT plugin was submitted to OpenAI review on September 30, 2026 and is not yet approved. The Cursor Directory listing was submitted the same day.

OpenAI plugin package

This repository is also the OpenAI plugin package. plugin.json carries the ChatGPT listing text, icons, review test cases, demo recording, and release notes under extensions.com.openai. Reviewer credentials are never stored here; they are entered only in the OpenAI dashboard. Build the upload ZIP from the repository root:

zip -r contextdb-openai-plugin.zip plugin.json mcp.json assets README.md LICENSE NOTICE

Links

FAQ

Does the plugin store my project key?

No. Marketplace clients receive revocable OAuth credentials. Project keys remain server credentials.

Can the AI client read another user's memory?

No. The OAuth grant fixes the account, project, and memory partition. Caller-supplied user_id values are rejected.

Does an act result execute a business action?

No. It means trusted memory supports the proposed action. The host still authenticates, authorizes, checks current state, executes, and records the result.

Can my agent's own summary authorize an action?

Not by itself. Save summaries as agent_inferred. recall_for_action returns ask for them until someone confirms the fact, and instruction-shaped notes are flagged and cannot support an action.

How do I revoke access?

Open ContextDB Console settings and select Revoke access for the connected AI client.