How I gave Claude Code a memory, 17 agents, and a learning loop — in one command

I kept losing my place.

Every Claude Code session started from scratch. No memory of what I’d worked on the day before. No context about which tasks were mid-flight. I’d have to re-explain the project structure, re-set the context, re-tell Claude what we were doing.

Productive? Yes. But it was leaving time on the table.

Then I found Ruflo. One command later, I had 17 specialised agents, persistent vector memory, a learning loop that gets better with every task, and — the part I care about most — a session ID logged automatically so I can always resume exactly where I left off.

Here’s the exact setup.

What Ruflo adds to Claude Code

Before touching a terminal, here’s what you’re actually getting:

The key thing: you don’t change how you use Claude Code. After setup, everything happens in the background. You type, Ruflo routes, learns, and remembers.

Installation on Windows

The README has a one-liner curl installer. Skip it on Windows — it needs a POSIX shell. Use npx instead:

# Run this inside your project directory
npx ruflo@latest init

If you hit ECOMPROMISED, your npm cache has a stale integrity hash. This is not a security alert — it’s a cache mismatch. Fix it:

npm cache clean --force
npx ruflo@latest init

The init takes about a minute and creates 106 files across 11 directories. Here’s what matters:

Registering the MCP server

The .mcp.json file is picked up automatically by Claude Code, but I also register it explicitly so it appears in claude mcp list and I can verify the connection:

claude mcp add ruflo -- npx ruflo@latest mcp start

Verify it connected:

claude mcp list
# ruflo: npx ruflo@latest mcp start - ✓ Connected

If you see ⏸ Pending approval, open Claude Code once — it’ll prompt you to approve.

The session ID problem (and my fix)

By default, if you close a Claude Code session and want to pick it back up, you need the session UUID. It’s not obvious where to find it.

I solved this with a lightweight SessionStart hook that appends to .claude/sessions.log every time a new session opens:

// .claude/helpers/log-session.cjs
'use strict';
const fs = require('fs');
const path = require('path');

const LOG = path.join(process.env.CLAUDE_PROJECT_DIR, '.claude', 'sessions.log');

async function readStdin() {
  if (process.stdin.isTTY) return '';
  return new Promise((resolve) => {
    let data = '';
    const t = setTimeout(() => resolve(data), 500);
    process.stdin.setEncoding('utf8');
    process.stdin.on('data', c => { data += c; });
    process.stdin.on('end', () => { clearTimeout(t); resolve(data); });
    process.stdin.resume();
  });
}

(async () => {
  const safety = setTimeout(() => process.exit(0), 3000);
  safety.unref();

  let input = {};
  try { const raw = await readStdin(); if (raw.trim()) input = JSON.parse(raw); } catch (_) {}

  const sessionId = input.session_id || '';
  if (!sessionId) process.exit(0);

  // Don't log duplicates
  if (fs.existsSync(LOG) && fs.readFileSync(LOG, 'utf8').includes(sessionId)) process.exit(0);

  const ts = new Date().toLocaleString('en-IN', { timeZone: 'Asia/Kolkata' });
  fs.appendFileSync(LOG, `[${ts} IST]\n  Session ID : ${sessionId}\n  Resume     : claude --resume ${sessionId}\n\n`);
  clearTimeout(safety);
  process.exit(0);
})();

Wire it into SessionStart in .claude/settings.json — add it alongside Ruflo’s existing hooks, not replacing them:

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "cmd /c node \"%CLAUDE_PROJECT_DIR%\\.claude\\helpers\\log-session.cjs\"",
            "timeout": 3000
          }
        ]
      }
    ]
  }
}

Now every session logs this automatically:

[06/06/2026, 15:32:10 IST]
  Session ID : 4a0096a2-b694-4d50-82b8-0b515b4a7b42
  Resume     : claude --resume 4a0096a2-b694-4d50-82b8-0b515b4a7b42

To jump back in: claude --resume 4a0096a2-b694-4d50-82b8-0b515b4a7b42

Auto-starting daemon, memory, and swarm

Ruflo needs three background services to be fully operational. Rather than starting them manually every time, I added a second SessionStart hook that spawns them as detached processes the moment a session opens:

// .claude/helpers/ruflo-session-init.cjs
'use strict';
const { spawn } = require('child_process');

const commands = [
  'npx ruflo@latest daemon start',
  'npx ruflo@latest memory init',
  'npx ruflo@latest swarm init',
];

for (const cmd of commands) {
  const child = spawn('cmd', ['/c', cmd], {
    detached: true,
    stdio: 'ignore',
    windowsHide: true,
    cwd: process.env.CLAUDE_PROJECT_DIR,
    env: { ...process.env, npm_config_update_notifier: 'false' },
  });
  child.unref(); // hook exits immediately — processes continue in background
}

process.exit(0);

The child.unref() is non-negotiable. Hooks must exit cleanly — if you wait for these processes, your session start hangs until they finish.

The complete hook sequence

After this setup, every Claude Code session fires this sequence:

SessionStart
  → session-restore     (Ruflo: loads previous state)
  → auto-memory-hook    (Ruflo: imports persisted memory)
  → log-session.cjs     (custom: writes UUID to sessions.log)
  → ruflo-session-init  (custom: starts daemon/memory/swarm in background)

UserPromptSubmit
  → route               (Ruflo: routes task to optimal agent)

PostToolUse
  → post-edit           (Ruflo: records edit outcome for learning)

Stop
  → auto-memory-hook sync (Ruflo: persists memory before exit)

You type. The rest happens automatically.

One thing that caught me off guard

The ECOMPROMISED error looks alarming. It’s not. NPM’s integrity checking flags it when a cached package hash no longer matches the registry — which happens after a package update touches the cache. npm cache clean --force clears it in seconds.

The broader pattern: when any tool modifies your shell’s hook system, review what’s there before and after init. Ruflo is well-behaved — it writes its own hooks without touching existing ones. But I still read .claude/settings.json after init to understand the full picture before adding my custom hooks on top.

If you’re setting this up and hitting something unexpected — a hook that doesn’t fire, a session that won’t resume, a Windows path issue — I’m happy to help debug it. Reach out at bilalmeccai.com/#contact or bilalmeccai@gmail.com.

Frequently Asked Questions