Meta Built It Themselves. That's the Problem.
A PM at Meta published an internal post earlier this year.
The title was: “I finally built my second brain.”
Not “I tried a new tool.” Not “this feature is useful.” A declarative. A before/after statement. The kind of language people use when something actually changed.
The system she was describing wasn’t a product. It was an internal experiment built by Meta’s analytics team — an AI agent with persistent, structured access to everything she was working on. Projects, meeting notes, commitments, context. Something that didn’t just answer questions. Something that tracked what was happening across her work and built on every prior conversation.
Within months, 60,000 people at Meta were using it.
What they built
The system Meta built does something specific.
It has memory. Not the 128K context window kind — the kind that persists. Episodic memory of past conversations. Semantic memory of what a person is working on. Procedural memory of how they work. When you talk to it, it already knows the situation. You don’t re-explain. You continue.
LinkedIn shipped something structurally similar in April. They called it the Cognitive Memory Agent — a shared memory substrate that sits underneath their AI systems. Memory that persists across agents, across sessions, across the full arc of what’s actually happening.
The framing at LinkedIn was about making AI agents smarter. But the actual achievement is different: agents that don’t make you re-explain the situation every time. Agents that already know what matters.
Both companies built this internally. Both built it because they needed it.
What they’re not saying out loud
There’s a fact embedded in both of these stories that doesn’t get named.
When Meta’s analytics org builds an internal AI second brain that spreads to 60,000 employees, it’s because no product existed that could do this. When LinkedIn engineers build a cognitive memory layer from scratch, it’s because the infrastructure they needed wasn’t available to buy.
These are some of the most well-resourced engineering organizations in the world. If this problem had an existing solution, they would have purchased it.
It didn’t. So they built it.
That’s not a commentary on Meta’s appetite for internal tools. It’s a market structure signal. The most obvious question is the one nobody is asking loudly: if Meta had to build this to get it — what does everyone else do?
The coordination gap at the individual level
The version Meta built works within Meta. LinkedIn’s version works for LinkedIn’s systems.
Neither was designed for the person who crosses ten different companies’ tools in a single workday. The person who received a commitment in Slack, needs to act on it in Jira, and knows it connects to something in email from three weeks ago — but can’t surface it fast enough to matter.
That gap isn’t technical. The technology clearly works. Meta validated it at 60,000-person scale. The gap is a product gap. The infrastructure exists inside big companies, for the people who work there, within the systems those companies control.
The knowledge worker who isn’t Meta doesn’t have access to this.
They’re still the coordination layer themselves. Still re-explaining context in every tool. Still manually remembering what they committed to, what they followed up on, what they said they’d do. Still losing time not to hard decisions — but to the work of holding things together.
This is what coordination failure looks like at the personal level. Not a missing feature. A missing layer.
What the experiment proved
Meta ran the most important test in the second-brain category this year. Not a case study. An internal product that 60,000 people adopted organically, starting with a PM who couldn’t stay silent about finally having the thing she’d been looking for.
The experiment answered the question the market keeps hedging: does this actually work? Is persistent, cross-context, commitment-aware AI useful enough to change how people work?
The data from inside Meta says yes.
Now the question is who builds it as a product — not just as an internal experiment. Who makes the version that works across tools, not just within one company’s stack. Who ships the thing the PM would have bought, if it had existed to buy.
Meta’s AI Second Brain experiment was published by Analytics at Meta in April 2026. LinkedIn’s Cognitive Memory Agent (CMA) was reported by InfoQ in April 2026.
