How a 30-person DTC brand cut new-hire ramp from 4 weeks to 3 days with Claude Code, open-source memory, and CLO

TL;DR — A 30-person DTC skincare brand ("Linea" — composite of three customers we've talked with) was burning ~60 hours of founder and senior-CX time on every new CX hire's first month. In one Friday afternoon they wired up a stack of three open-source tools — Claude Code, obsidian-mcp-tools, and Cognition's CLO — and dropped their next hire's ramp time from 4 weeks to 3 days. Total setup cost: $0 in software, 45 minutes of one ops person's time. This post is the exact walkthrough.

---

The problem every founder-led company has

Linea has 30 people. Eight of them are in customer experience. The founder, Maya, personally wrote every SOP, every escalation path, every tone-of-voice doc, every "here's why we refund without asking" rule. All of it lives in Notion, Slack, a Google Doc from 2024, and Maya's head.

When a new CX rep joins, the pattern is always the same:

1. Week 1 — they read docs. Half are outdated. They don't know which half.
2. Week 2 — they shadow. They ask the same question 11 times across three seniors.
3. Week 3 — they start taking tickets. They ping senior staff every 7 minutes.
4. Week 4 — they're productive-ish.

Maya does the math one day. Four weeks × $22/hour × 40 hours + ~10 hours of senior-CX time shadowing + ~4 hours of her own time per hire × 6 hires/year = ~$32,000 a year, most of it invisible.

She doesn't need a fancy LMS. She doesn't need a course builder. She needs her own brain available to new hires the moment they log in, and she needs to know — weeks later — what they actually kept.

That's the job of this stack.

The stack in plain English

Three parts. Each part is a thing a single person can install. Together they form an AI operating stack.

1. The agent — Claude Code (or Claude Desktop, or Cursor — any MCP-speaking client). This is the thing the new hire actually talks to.
2. The memory — an organizational memory server that holds the company's docs, SOPs, prior tickets, and tribal knowledge. For Linea: obsidian-mcp-tools, because they already wrote everything in Obsidian.
3. The pedagogy layer — Cognition CLO. Sits on top of the memory stack, models each employee's retention over time using the Weibull forgetting curve, and pings them before they forget.

That's it. The agent answers. The memory remembers. CLO is the chief learning officer Linea never had the headcount to hire.

Step 1 — Pick your memory engine (5 minutes on /memory)

Maya opens cognitionus.com/memory. She's not a developer. She sees 19 curated open-source tools, grouped by who each tool is for:

• No-code · Obsidian as your brain — for teams who already keep notes.
• Self-building knowledge bases (the Karpathy pattern) — the frontier, for teams with a technical lead.
• MCP-native memory servers — one-line drop-ins for any agent.
• Full agent memory frameworks — production-grade infra.

She clicks the card for obsidian-mcp-tools. The plain-English explanation tells her:

If your team already writes notes in Obsidian, this plugin makes those notes readable and editable by Claude or Cursor. Your agent can answer questions from your notes, add new ones, and link them together — so your meeting notes, SOPs, and internal docs become your agent's long-term memory. No code required.

She sees setup time: 5 minutes. She sees good for: non-technical founders with an Obsidian vault in use. That's her. She clicks Open repo, follows the install-from-community-plugins instructions, done in 4 minutes.

What just happened: Linea's Obsidian vault — 186 internal docs — is now readable by Claude Code. Any employee pointed at this vault via .mcp.json gets Maya's brain as part of their agent's working memory.

Step 2 — Build the rest of the stack (8 minutes on /stack)

Same session. Maya clicks Build my stack in 5 steps. She types:

"30-person DTC skincare brand. 8-person CX team. New CX reps ramp in 4 weeks — I want that down to 1 week. We already use Slack, Shopify, Gorgias, Notion."

Gemini spits back a recommendation tagged as the e-commerce / DTC archetype:

• claude-code — the personal agent
• obsidian-mcp-tools — her already-picked memory
• gorgias-mcp — so the agent can read live ticket history
• slack-mcp — so the agent can read team threads
• cognition-clo — the pedagogy layer

Each step is a card. She can accept, tweak, or skip. She accepts.

At the end, the site generates her .mcp.json:

{
  "mcpServers": {
    "obsidian-mcp-tools": {
      "command": "npx",
      "args": ["-y", "obsidian-mcp-tools"]
    },
    "gorgias-mcp": {
      "command": "npx",
      "args": ["-y", "@gorgias/mcp"]
    },
    "slack-mcp": {
      "command": "npx",
      "args": ["-y", "@slack/mcp"]
    },
    "cognition-clo": {
      "command": "npx",
      "args": ["@cognition/clo-mcp"],
      "env": { "COGNITION_API_KEY": "cog_live_…" }
    }
  }
}

She pastes it into .claude/mcp.json in a shared config repo. Every CX rep who clones the repo gets the whole stack.

Step 3 — The 7-day playbook (1 click, rate-limited, free)

Below the mcp.json, the site generates a personalized 7-day onboarding playbook for Linea's specific stack. Not generic advice — actual actions:

| Day | Action | |-----|--------| | Mon | New hire runs claude mcp list, confirms all 4 servers connect. Reads Maya's "why we exist" note first. | | Tue | Agent-led tour of last 50 Gorgias tickets: "classify each one by escalation type." | | Wed | Hire drafts 10 replies in Claude; Maya reviews the first one live, then the agent reviews the next 9. | | Thu | Claude + memory generates a personalized cheatsheet from the hire's 10 drafts — the 3 patterns they're weakest on. | | Fri | Shadow 2 live tickets with senior CX. Agent watches (MCP → Gorgias). Writes a postmortem. | | Sat | Off. CLO's Weibull model schedules a 48-hour review nudge for Monday. | | Mon | CLO pings the hire with a 4-question quiz on the concepts they're about to forget. Two minutes, done. |

Plus 5 daily habits and 4 copy-pasteable starter prompts. All rate-limited (5 generations per IP per hour) because every hit costs real money on Gemini.

Step 4 — Ship it. Measure what they forget.

The hire (we'll call her Amara) starts on Monday. She opens Claude Code. Her first prompt:

"You're my onboarding partner for Linea's CX team. Walk me through my day-one plan from the playbook. Assume I'm new."

Claude pulls from Obsidian (Maya's SOPs), Gorgias (real ticket history), and Slack (team threads). It answers like a senior teammate would, but patient and available at 2am if she's stuck.

Meanwhile, Cognition CLO is silently logging every concept Amara engages with — tone rules, escalation triggers, refund policy corners — and applying a Weibull survival curve per concept.

On day 3, CLO's dashboard shows Maya exactly which concepts have dropped below the 60% retention line. Maya doesn't have to guess. She doesn't have to test. She gets a one-line notification: "Amara is 72 hours away from forgetting how to handle chargeback disputes — send her the 3-minute primer."

She sends it. Two minutes of her day. $0 in software.

The result

• Ramp time: 4 weeks → 3 days to first solo shift, 2 weeks to full productivity.
• Senior-CX time saved per hire: ~14 hours.
• Maya's time saved per hire: ~3.5 hours.
• Total cost of the stack: $0 for the open-source layer, CLO priced per seat.

The stack doesn't replace training. It replaces the repeated part of training — the "why do we refund without asking" answer that Maya has typed into Slack 40 times. Now she types it once, into Obsidian. Every future hire gets it the moment they need it.

What this unlocks beyond onboarding

Onboarding is the sharp end. The same stack handles:

• SOP decay: CLO tracks which rules the whole team is drifting from.
• Post-incident review: after any big customer mess, write the postmortem in Obsidian. CLO schedules it as a refresher across the team three weeks later, not a one-time Loom.
• Vendor churn: when a new 3PL replaces the old one, the agent rewrites affected SOPs. CLO tests the team on the new rules without Maya playing Jeopardy host in Slack.

This is what "organizational memory" actually means as a product category. Not storage. Not search. Knowing what your team knows, and what they're about to forget.

---

FAQ

Is this only for Claude Code users?

No. Any MCP-speaking client works — Claude Desktop, Cursor, Aider, Zed, Gemini CLI, OpenCode. The registry tools and CLO are all MCP-native.

Do I need a developer to set this up?

For the obsidian-mcp-tools path Maya took: no. Install from the Obsidian community plugin store. For other paths (gbrain, memory-mcp, mem0) you'll want someone technical for 30–45 minutes of setup. The registry labels each tool's difficulty.

Does my org data leave my servers?

The registry tools all self-host. CLO never stores your source documents — it models retention from concept events (what was studied, when, by whom) and optional spans, not from your SOPs themselves. Your IP stays in your repo.

What's CLO actually doing under the hood?

Weibull survival curve per employee per concept: R(t) = exp(-(t/S)^β). Two parameters fit per concept per person from their engagement history. The model tells us when each concept will drop below a safety threshold. We send the nudge before it does.

How do I get started in under 15 minutes?

1. Go to cognitionus.com/memory — pick your memory tool.
2. Go to cognitionus.com/stack — describe your company, accept the stack.
3. Paste the generated .mcp.json into your .claude/ folder.
4. Add CLO's API key.

That's it. The first time a hire logs in, the stack is waiting for them.

I already use Notion, not Obsidian. Does this still work?

Yes. Swap obsidian-mcp-tools for Notion's MCP server. The pattern is identical.

---

Want the same walkthrough tailored to your company? The /stack builder is free. Email vedant@cognitionus.com if you want us to turn CLO on for your team — we'll wire it in ~30 minutes, no demo call required unless you want one.