How to give Claude Code and Codex persistent memory over MCP

By now you've probably felt the pain: your coding agent is great within a session and amnesiac across sessions. The fix is a persistent memory layer your agents can reach over the Model Context Protocol. This walkthrough uses Knownbase, a hosted MCP memory server, but the shape is the same for any MCP-compatible tool.

Step 1: Create a workspace and an API key

Sign up for a free workspace (no card required). Once you're in, open Settings → Workspace & MCP and generate an API key. It'll look like kb_.... Copy it now. For security the full key is shown only once.

A workspace is your memory boundary. Every note you store is scoped to it and isolated from every other tenant, so a personal project and a client project can each have their own key.

Step 2: Add the MCP server to your agent

Most MCP clients read a JSON config that maps a server name to a URL and headers. Point it at the Knownbase endpoint and pass your key as a bearer token:

{
  "mcpServers": {
    "knownbase": {
      "url": "https://knownbase.dev/mcp",
      "headers": {
        "Authorization": "Bearer kb_YOUR_KEY_HERE"
      }
    }
  }
}

Claude Code can add it from the command line instead of editing config by hand:

claude mcp add --transport http knownbase https://knownbase.dev/mcp \
  --header "Authorization: Bearer kb_YOUR_KEY_HERE"

Codex, Cursor, ChatGPT, and other clients take the same URL, key, and header. Drop them into wherever that client configures MCP servers. Restart the agent so it picks up the new server.

Step 3: Confirm the tools are live

After connecting, your agent should see a set of Knownbase tools. The ones you'll use most:

• upsert_note: create or update a note (title, project, body, tags, status).
• search_notes: full-text and semantic search across the workspace.
• get_note / get_notes: fetch full content by id, singly or in batches.
• list_projects, list_note_revisions, workspace_info: browse and audit.

Ask your agent to "list the projects in Knownbase." An empty result means it's connected and working; you just don't have any notes yet.

Step 4: Store something worth remembering

Now put it to use. At the end of a working session, ask the agent to save a handoff:

"Save a handoff note to Knownbase in project billing summarizing what we decided about the webhook retry logic and what to do next."

Behind the scenes the agent calls upsert_note with a title, the billing project, and a markdown body. The note is now durable: it survives this session ending. (Not sure what to put in a handoff? See Agent handoffs.)

Step 5: Recall it in the next session

Tomorrow, in a fresh session (even in a different agent), start with:

"Search Knownbase for what we decided about billing webhooks before we continue."

The agent calls search_notes, pulls the relevant note, and picks up with full context instead of re-deriving it. That's the whole loop: store as you go, recall on demand.

Tips for getting the most out of it

• Use projects and statuses. Organizing notes into projects with active/todo/done/blocked statuses keeps search results sharp.
• Link related notes. A [[Note title]] reference in a body becomes a real backlink, so a handoff can point at the decision it depends on.
• Add it to your agent's instructions. A line like "save decisions and handoffs to Knownbase, and search it before starting" makes the habit automatic.
• Share a workspace with teammates. On Team and Business plans, everyone's agents read and write the same memory.