Claude Code + deja

Why Claude Code + deja?

Claude Code is powerful out of the box, but it forgets everything between sessions. Every time you start a new conversation, you re-explain the same deployment quirks, the same architectural decisions, the same “don’t touch that file” warnings.

With deja, Claude Code builds a persistent memory layer. When it learns that your production deploy requires a specific wrangler flag, it remembers. When it discovers that your monorepo has a tricky circular dependency, it stores that. Next session, before touching anything, it injects the relevant memories and already knows what to watch out for.

Prerequisites

• A deployed deja instance (Cloudflare Worker URL)
• Your deja API key (Bearer token)
• Claude Code (via Claude Desktop or CLI)

Setup

Option 1: MCP (Recommended)

The fastest path. Add deja as an MCP server so Claude Code can call learn and inject as native tools.

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "deja": {
      "url": "https://deja.your-subdomain.workers.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_DEJA_API_KEY"
      }
    }
  }
}

Claude Code CLI — edit .claude/settings.json in your project root:

{
  "mcpServers": {
    "deja": {
      "url": "https://deja.your-subdomain.workers.dev/mcp",
      "headers": {
        "Authorization": "Bearer YOUR_DEJA_API_KEY"
      }
    }
  }
}

After saving, restart Claude. You should see deja listed in the available MCP tools. Claude Code can now call learn and inject directly.

Option 2: System Prompt Approach

If you prefer to guide Claude Code through REST calls, add instructions to your system prompt or CLAUDE.md:

## Memory (deja)

Before starting any task, recall relevant memories:

    curl -s -X POST https://deja.your-subdomain.workers.dev/inject \
      -H "Authorization: Bearer $DEJA_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"context": "<describe current task>", "scopes": ["shared"]}'

When you learn something important (a gotcha, a pattern, a decision), store it:

    curl -s -X POST https://deja.your-subdomain.workers.dev/learn \
      -H "Authorization: Bearer $DEJA_API_KEY" \
      -H "Content-Type: application/json" \
      -d '{"trigger": "<when this applies>", "learning": "<what to remember>", "scope": "shared"}'

Option 3: deja-client in Scripts

Use deja-client in helper scripts that Claude Code can execute:

npm install deja-client

import deja from "deja-client";

const mem = deja("https://deja.your-subdomain.workers.dev", {
  apiKey: process.env.DEJA_API_KEY,
});

// Before a task — inject context
const { prompt, learnings } = await mem.inject("deploying to production");
console.log(`Recalled ${learnings.length} relevant memories`);
console.log(prompt);

// After a discovery — store it
await mem.learn(
  "deploying to production",
  "always run wrangler deploy --dry-run first — the binding names changed in v3",
  { scope: "deploy", source: "claude-code" }
);

Example: Deployment Memory

Here is a real workflow showing how Claude Code builds deployment knowledge over time.

Session 1 — Claude Code deploys and hits an issue:

# Claude Code runs the deploy and it fails
$ wrangler deploy
# Error: binding "AI" is not defined in wrangler.toml

# Claude Code fixes it, then learns from the experience

Claude Code calls learn (via MCP or curl):

{
  "trigger": "deploying to production with wrangler",
  "learning": "The AI binding must be explicitly declared in wrangler.toml under [ai]. Without it, the worker crashes on cold start. Add: [ai]\nbinding = \"AI\"",
  "scope": "deploy",
  "confidence": 0.95,
  "source": "claude-code",
  "reason": "Deploy failed with binding error — this is easy to forget after adding new features that use the AI binding"
}

Session 2 — Next time Claude Code is asked to deploy:

Claude Code calls inject with context "deploying to production with wrangler" and gets back:

Relevant memories:
- The AI binding must be explicitly declared in wrangler.toml under [ai]. Without it, the worker crashes on cold start.

Claude Code checks wrangler.toml before deploying. No more forgotten bindings.

What to Learn, What to Inject

Related

• Cursor + deja — same MCP approach, different editor
• MCP Filesystem + deja — combine file awareness with memory
• Deployment Memory pattern — structured approach to deploy learnings