Everyone is talking about Karpathy's LLM knowledge base. 16 million views in a few weeks. But nobody is naming the actual problem it reveals.
Karpathy built a personal wiki, roughly 100 articles and 400,000 words, compiled entirely by an AI from his raw research docs. He didn't write a single word of it. It's one of the most compelling demos of what personal AI knowledge infrastructure can look like.
But here's the problem: The wiki lives on his machine. Every time he opens a different AI tool, it knows nothing about it.
That's not a Karpathy problem. That's everyone's problem.
Some other people like Allie Miller built a "context vault," structured summary docs she feeds into Claude to avoid re-explaining herself every session. Teams are using Notion as a memory layer. Developers are writing custom scripts to sync context across tools. Some are setting up github wikis, while some are using Obsidian to set up local markdown files.
All brilliant. All manual. All solving the same root failure: AI has no durable, portable sense of who you are.
Every time you open Claude, Cursor, ChatGPT, or whatever ships next month, it starts from zero. There is no standard for "this is who I am" that travels with you. The burden of continuity falls entirely on the user.
Some will point to Claude's memory feature or ChatGPT's memory. These are genuine steps forward. But they are per-platform, opaque, and you don't own them. Switch tools and you start over. They are not solving the
portability problem. They are each building their own silo of you.
The real issue is that context continuity is being treated as a feature when it is actually infrastructure.
Features get built per-product. Infrastructure sits beneath products. What's missing is a memory layer that sits above any individual tool. One canonical store of your context, preferences, and knowledge that you own, control, and selectively grant any AI permission to read. Not stored by Anthropic. Not stored by OpenAI. Portable across the fragmented AI landscape.
A year ago the coordination problem here would have been enormous. Getting AI providers to query an external context layer would have required deals, standards negotiations, and a lot of waiting.
MCP changed that. MCP (Model Context Protocol) is an open standard that lets AI tools query external data sources directly. Any AI tool that supports MCP can already read from an external context server, no special agreements with providers required. The infrastructure exists. What has been missing is someone building the right thing on top of it: a user-owned, canonical context store that any MCP-compatible AI can read with your permission. For example, with the AICF MCP Server, you could not only read but directly add, edit, delete, organize your cross-platform memories from within the AI chats.
Our vision: One memory. Every AI. Owned by you.
p.s. we have a lifetime deal going on AppSumo which will end this month: https://appsumo.8odi.net/m4n0da

Hard to believe now, but back in 2024 when we named our concept the "Open Context Layer," every pitch and networking event ended the same way:
"What the heck is context? Who needs it? Why would I even bother?"

I'd spend 3-5 minutes explaining that context is your data, in a portable format, pluggable into any platform. I cycled through every analogy I could find: "think of it like a USB stick," "imagine carrying a briefcase with all your important documents," "a Google Drive that connects to platforms."

But now when I say "we're building an open context layer" ... people just get it.

We're quick to complain about the hard parts of startup life. But sometimes there is a silver lining..

Most memory tools I have tried solve storage and retrieval well, but still rely on me to manually save context. The part that kept breaking for me was capture: the useful context is the call I just had or the thing I was reading 20 minutes ago, and I never remember to save it in the moment.
We took a shot at the ambient version of this with Minimi. It captures screen activity and call audio on your Mac, keeps it encrypted on-device, and exposes it back to Claude over MCP, so retrieval happens inside the chat you are already in.
Curious how this community thinks about the manual-save vs ambient-capture tradeoff, and whether on-device changes the calculus for you.
If you want to see the approach, it is on our PH page from this week and I would value honest feedback.

The numbers tell the story:

3264 users

4.99 rating

5.3% paid conversion

Our vision has been the same from day one: one knowledge base that works across all your AI tools.
And to achieve that, we launched our MCP Server in February. It connected to Claude Desktop, Claude Code, LM Studio, and most major AI agents. As far as we were concerned, the MCP chapter was closed.
Then we launched on AppSumo.
Within days, users kept asking for one thing: " ?"
Then came the questions about headless agents. CI scripts. n8n workflows. Users had intricate setups. They wanted a memory store for their OpenClaw agents, which they could also plug into Claude, which they could also call from a workflow runner. One memory, three completely different environments.
That's when we realized our MCP Server had a problem: it only supported OAuth. *facepalm*
(Getting a bit technical here, bear with me)
OAuth assumes there's a user sitting at a UI who can click "sign in" in a browser window. That's fine for Claude Desktop. It falls apart the moment you're running a headless agent on a server, or chaining four tools together in an automated workflow. Nobody is there to click anything!
So the team got to work. A few days later, we shipped Personal Access Tokens (PATs) for the MCP Server.
And that's how we ended up being the that works in three places at once:
In your browser, as a sidebar
Inside your chat agents, as an extension or MCP server
In your programmatic workflows, as an MCP with PAT-based authentication
New setup guides for everything are at docs.plurality.network.
Maybe a weekend project: give your OpenClaw or n8n agents a memory. Make them less forgetful, more intelligent, and a lot more useful.
If you are already running such setups, we have a lifetime deal going on for AI Context Flow: https://appsumo.8odi.net/m4n0da

box . ' .
( _ )

______________________________ .
Imagine: You're deep into a technical doc. Three browser tabs open. A Wikipedia article on one, an API reference on another, a blog post explaining a workaround on the third.

You find exactly what you need.

But first, what is MCP?
MCP Servers, often called the USB-C of AI systems, have had a meteoric rise in 2025.
And the reason is plain to see. Without MCP, the complexity of integrating AI with external systems rises quadratically (esp. with the proliferation of AI agents). With MCP, it only increases linearly.
Practically, what this means is that you can connect your AI agents with any external data/tool source that exposes its functionality as per the standards of MCP. Kind of like a REST API, if you are an oldie like me in tech.
:
The MCP Servers I have been using the most (will create separate posts for these in the coming days):
: I tell ChatGPT what I want, and then use the Canva MCP to create an editable image of my concept. The file gets automatically added to my Canva account and I fix any issues manually. We all know AI-generated images have weird problems (esp. with spellings etc.), so Nano Banana wasn't really useful for me. A lot of my posts have pictures created this way (ping me if you want me to put together a short tutorial on this).
: Same as above. I use it to create mind maps of my brainstorming sessions with AI, where I can (sort of) create decision trees of how we arrived at a certain point.
: I haven't done it yet. I'm scared for privacy reasons, but the urge is strong. I hate meeting management, and as a founder, mother, and IT professional, I have to maintain my schedule down to the minute if I want to balance the various areas of my life.
: I have a lot of my context in Plurality's Memory Studio. With the Plurality MCP, I don't have to use the AI Context Flow browser extension and can instead use the MCP server directly on Claude Desktop. This helps me in three ways:
  I can pull in my context on Claude Desktop, Claude Code, and even with Cowork
  I don't have to keep pressing the optimize button to pull in context (MCP is seamless. Just chat, and it will do the rest)
  The overall experience is muuuuccchhh better than the extension

This guide shows how to add Plurality (or any other MCP server) to your different agents.
Which MCP servers are you using the most? Let's exchange notes.

We built AI Context Flow because we believe the future of AI is about making all your tools and knowledge work as one. Using one AI agent has been the past and multi-agent is the future.

And every time you switch, your context should come with you. This is for you if:

• Your AI agents lose context every time you switch between projects

• You are drowning in saved articles that never talk to your AI

• You re in a team where everyone uses different tools but nobody shares the same context

We are offering $59 for lifetime access available for a limited time. You were the first ones to discover us, so it s only fair that you get the first lifetime plans.

: '

Ever asked ChatGPT about "the login system" and watched it confidently mix up three different projects?

That's memory bleed i.e. when your AI assistant can't separate your contexts. Your healthcare client's specs bleed into your fintech code.

Your Python project gets Node.js suggestions from yesterday's conversation. Your personal writing style contaminates professional emails.

700,000 people have quit ChatGPT in the past few weeks. Political backlash, ethical concerns over ICE using GPT-4, and a drop in product quality are all driving a massive wave of switching.

Most are landing on Claude. Some on Gemini. Many on both.

But every single one hits the same wall: how do you move your context?

All the preferences, projects, and workflows built up over months or years stay locked inside the old platform. The new one has no idea who you are.