Why we're building a memory layer for organizations

There's a problem nobody has named yet.

Every company runs on what its people know — the procedure, the protocol, the unwritten reason a Q3 pricing experiment failed and shouldn't be tried again. That knowledge fades constantly: about 23% of it per quarter on the curve, more if the team turns over. You don't notice it happening because there's nothing measuring it. You notice the consequences — the rework, the contradicted decisions, the customer who got told two different things by two different people on the same day.

We think there should be a fourth place where institutional knowledge lives. Not a wiki (nobody updates them). Not a channel (nobody reads them). Not a head (it walks out the door). A memory layer that's structurally part of the organization — that watches what gets learned, predicts what's about to be forgotten, and intervenes before the consequence shows up.

That's what Cognition is building.

Two products, one substrate

We're a research lab. We ship two products on the same memory layer.

CLO is what you can use today. It's the second brain that sits on top of your Claude Code, watches what you decide, models how you think over time, and catches contradictions before you act on them. "Heads up — you decided not to do this last quarter and that worked out. What changed?" That's the killer feature: not retention dashboards, not training compliance, but the moment when CLO catches you about to repeat a decision you already made and forgot about.

Personas is what we're building next, for a different buyer. Agent companies — the ones building coding agents, customer-support agents, tutor agents — need to test whether their agent serves a real human user, not a hypothetical one. The Personas marketplace will let them query a real founder's brain via API: their decisions, their reasoning patterns, what they remember and what they don't. Train and test against real personas, not toy benchmarks. Coming soon, waitlist open at /personas.

Both products are powered by the same primitive: a model of how a specific person learns, decides, and forgets, derived from a forgetting-curve framework with consolidation dynamics on top. We're not building a chatbot or a wiki. We're building a memory layer that AI agents can read from and write into.

Why now

Three things converged.

One: agents are at the point where they're about to make consequential decisions on behalf of humans. A customer-support agent that recommends the wrong return policy because it doesn't know which policy this specific customer was promised yesterday is a real liability. The agent needs the customer's memory model, not just their transcript.

Two: the open-source memory engine ecosystem matured. mem0, letta, zep, gbrain, supermemory, basic-memory — there are a dozen credible substrates for storing and retrieving context. None of them model forgetting. They store what you tell them; they don't tell you when you should be reminded. We sit on top of all of them.

Three: Anthropic shipped MCP. We don't have to integrate with everything in your stack — we sit underneath your Claude Code, which is already integrated. Every CC session is a write to the memory layer. Every CC session is a read from it. The bidirectional loop is finally cheap.

What we're not building

We're not building a wiki. The wiki is dead. We're not building a chatbot — there are 50. We're not building Notion or Confluence or Granola or Reflect. Those are surfaces. We're the substrate.

We're not building a generic-LLM tool. We're built on Claude Code via MCP. That's a structural bet on the agent layer Anthropic is building. We'll add Cursor, Codex, and other MCP-spec compliant agents as they become first-class, but we lead with CC.

We're not building Personas yet. Personas needs supply (1000+ users with real brains), demand (agent companies who pay to query), and a consent layer that's defensible. CLO is the wedge that produces supply. We'll launch Personas when supply makes the marketplace credible. Until then it's a waitlist.

How to follow along

This blog is where we publish what we learn. Eval results from the memory marketplace. Patterns we see in real CEO brains (anonymized). Failures and what we learned. The lab claim is earned through public learning, not asserted.

If you want to use CLO today: cognitionus.com/onboarding/welcome — five minutes, $199/mo, on Claude Code.

If you're building agents and want early access to Personas: cognitionus.com/personas.

If you want to read what we ship: keep an eye on this page. We aim for one post a week.

— Vedan, Joseph, and the team