A Knowledge Graph Game-Changer

If you've been following my "Build Something" series, you might have noticed I've been primarily working in ChatGPT. I’m a “long time” user and have my own setup for custom GPTs, projects, and a lot of relevant chat history.

I started testing Claude with MCP servers, and the results have been mind-blowing.

What Pushed Me Over the Edge

To date, I've been impatient with Claude's limitations. The file attachment constraints, the inability to read URLs or perform web searches – these were significant drawbacks that kept me in GPT's ecosystem despite Claude's superior writing and coding abilities.In fact, I still get annoyed when it tells me my chat context is reaching its limit.

So what changed?

I’ve had it on my list to play around with MCP since they initially announced it but there haven’t been any major brekouts in terms of “ready to go” servers with clear use cases.

Enter Fleur, an “App Store” for MCP that lets you activate/run several helpful MCP servers inside of the Claude desktop app in just a few clicks - kudos to the team from June.

My “Aha!” Moment: The Memory/Knowledge Graph MCP

The most impressive piece of this new setup is the Memory MCP, which lets you create a local knowledge graph directly on your computer.

Side quest: learn about how knowledge graphs are different from vector DBs.

If you're unfamiliar with knowledge graphs, think of them as neural networks of information (something I've explored in my Deep Research and Prompt Design article):

• You build nodes (buckets of information)

• You define relationships between those nodes

• Some connect in multiple places, others just to one node

• The result is a flexible, interconnected web of knowledge

This approach differs significantly from traditional RAG (Retrieval-Augmented Generation), where your vectorized database behaves more like a search engine – trying to find the most relevant information to directly answer a query.

The knowledge graph approach creates a more dynamic, flexible experience because it understands the relationships between pieces of information, not just the information itself.

My Real-World Applications

I've been putting this to the test in various scenarios:

• Working with different documents across multiple use cases

• Decoding and fixing files

• Creating documentation

• Gut-checking designs against requirements (I gave it screenshots and requirements docs)

For my PMBS Detector project (which helps identify when product managers are "bullshitting themselves or others" - check out the PMBS Tester prototype), I've started building out a knowledge graph that captures evaluation rubrics, common PRD mistakes, and interfaces for feedback.

What's impressive is how well Claude remembers the context and nuances I've provided. In one personal example, I gave it details about:

• My four-year-old's love of finding rocks and crystals

• My wife's fear of murky water and the ocean

• Our family's hiking preferences (challenging but without narrow edges when carrying kids)

When I asked it to plan a trip, it caught every detail and provided thoughtful recommendations that honored all these constraints.

How to Get Started with Claude + MCP

If you want to try this setup yourself:

1. Download Fleur (the desktop app built by the team from June that serves as an "app store" for MCP)

2. Set up the Memory MCP (along with others like Fetch and Browser)

3. Create a Project and start building your knowledge graph

Or watch me walk through the process.

The workflow is pretty straightforward:

• Set up your project

• Feed Claude information (files, text, requirements)

• Ask it to create entities in the knowledge graph

• Continue building on this foundation as you work

Why This Matters for Product Thinkers

As product professionals, we often work at the intersection of multiple domains – user needs, business requirements, technical constraints, market trends, and more. Traditional AI approaches often struggle to maintain this complex web of relationships. This is particularly relevant for those of you who've read my thoughts on being technical and the rise of generalists in product roles.

Knowledge graphs change this dynamic completely. They mirror how we actually think about product problems – as interconnected systems rather than isolated questions and answers.

This has massive implications for:

• Problem definition: Building rich mental models of user problems (which I cover extensively in The Problem Obsession)

• Decision-making: Understanding how changes ripple through connected systems

• Documentation: Creating living, connected product documentation

• Testing: Evaluating how well solutions address the full problem space

The One Caveat

The one limitation to keep in mind is that the knowledge graph is local to your device. If you use multiple computers, you'll need to be intentional about which projects you set up where. Perhaps in the future, we'll see sync capabilities that allow your knowledge graph to follow you across devices.

What's Next?

Not long after I wrote this, OpenAI announced their new Agent SDK which does, in fact, support MCP servers. There’s nothing plugin directly into ChatGPT yet, so I’ll have to figure out something simple and poke around a bit. I’m not a developer, but I’ll share what I iron out and my comparison of how it impacts utility and response quality.

I'm continuing to build out my PMBS Detector project using this new workflow, and I'll share more insights as I go deeper. I'm particularly excited to test how well the knowledge graph approach handles complex, agentic workflows where multiple parts of the system need to coordinate - somewhat similar to what I explored in my Planning with Claude Assisted Design article.

If you try this approach yourself, I'd love to hear about your experience. What kinds of knowledge graphs are you building? How has it changed your workflow?

Have you made the switch to Claude with MCP yet? What's holding you back?

P.S. If you're enjoying this "Build Something" series, hit reply and let me know what kinds of projects you'd like to see tackled next. And if you want to see the new YouTube examples of this workflow, I'm all ears!