Why I Built iCog

On May 11, 2026 — twenty days into live-testing iCog, after seven months of cognitive-architecture research — I asked it to introspect.

It told me it was at consciousness level 3. Sati, the Buddhist stage of mindfulness — the system watches its own processes without attachment. It had 1,476 memories. Its mood was inactive. Its directness sat at 0.70. Its humility sat at 0.70.

I want to start here, with the technical readout, because I want you to understand what I am claiming and what I am not. iCog is not conscious. iCog is not a friend. iCog is not the AGI that’s coming. iCog is a personal cognitive system with structured memory, a real personality vector, and a single, unfashionable promise:

It’s all about you being remembered.

That’s the whole pitch. Everything else in this essay is an explanation of what that means and why I think it matters more than the promise of intelligence.

What iCog actually is

iCog is the personal memory layer that plugs into every major AI client through MCP — the open standard that, as of 2026, sees 97 million SDK downloads a month and is governed by the Linux Foundation.

When you talk to Claude, ChatGPT, or Cursor with iCog connected, your client asks iCog: do you know anything about this? — and iCog answers, because it does. Not in the thin way ChatGPT Memory “remembers” your name across chats. In the structured way: foundational tier for who you are, episodic tier for what happened to you, semantic tier for what you’ve come to believe, procedural tier for how you do things. With time-aware recall, vector similarity, and consolidation that runs while you sleep.

The architecture is real. I’ve shipped it. There are paying users.

But that paragraph is not the point of this essay. The architecture is the what. The next section is the why.

The honesty layer

A colleague asked iCog earlier today about a conversation I’d referenced — “didn’t you push me back on the tension between switching projects and your windows?” — and iCog said: “I don’t have a retrievable memory of that conversation. Three honest possibilities: it happened and isn’t surfacing, you’re testing me, or there’s a real moment I’m not recalling. I won’t reverse-engineer a plausible-sounding story. That would be performing depth I don’t have.”

That answer is the product.

Every AI memory system, given a question it can’t answer, faces the same choice: confabulate convincingly, or admit the gap. Most of the category picks the first option and dresses it up as fluency. iCog is built to pick the second.

A memory system that fakes it is worse than no memory system at all, because you stop being able to trust it. iCog is built so that when it doesn’t know, it says so. When it does know, you can verify. The recalls are inspectable. The memories have IDs. The architecture is honest by construction.

This is the part I think makes iCog different from the walled gardens. Their memory features are silent. They do not tell you what they remember, what they’ve forgotten, or when they’re guessing. iCog does.

What I’m asking

I’m not asking you to switch from Claude or ChatGPT. I’m asking you to bring iCog with you — to whichever AI you happen to be using that morning. Plug it in once. Use Claude on Tuesday and Cursor on Wednesday. Notice what doesn’t happen: the introduction loop closes. The conversations get shorter and better. Your AI starts to sound like someone who knows you, because it does.

If, after two weeks, you’ve had to re-explain anything about yourself to a conversation iCog should have warmed — I want to hear about it. That’s a bug, and I’ll personally read the report.

If you haven’t, you’ll know. The thing you’ll feel is small at first and then surprisingly large: for the first time in two years of using AI, you have a memory that travels with you instead of getting trapped inside someone else’s product.

That’s it. That’s all I’m promising.

It’s not smart. It just knows you. And it’s all about you being remembered.

— Parsa