A company has always had two kinds of skill. There is the skill it pays for, which arrives with employees, consultants and contractors and leaves when they do. And there is the skill it owns, the know-how that is genuinely the firm's own and is usually what its advantage rests on.

Something has quietly appeared alongside them, and it does not fit either category.

When someone works on a real problem at a real company using AI, correcting the model, throwing out its first three answers, working out what actually succeeds on this particular codebase or account or regulator, they produce a third thing. Call it earned skill. It does not belong to the worker, because it is useless anywhere else. It does not belong to the firm in any recorded way, because nobody wrote it down. It is earned, on the spot, by doing the work.

Glen is a bet that this third kind of skill is about to become the most valuable thing a company produces, and that right now every company throws it away by the hour.

What they built

Glen is a memory layer for AI agents. It brings together agent sessions from tools like Claude Code, Codex and Cursor with the company's own material from Slack, GitHub, tickets and documentation. When an agent starts a new task, Glen finds what was learned before and hands it over.

Around that sit code review with the original background attached, search across what agents have done before, a shared library of skills, several agents working together, and the freedom to switch AI provider. The company reports roughly 29 per cent lower costs and tasks finishing 21 per cent faster.

It is based in San Francisco. Nikos Dritsakos, the founder, was previously Head of Special Projects at Composio and VP of Technology at FliteHouse, and bootstrapped his first company, SalesBop, to a $500,000 acquisition in eighteen months.

Who else is in this field

Giving models background information is a crowded business. Pinecone, Weaviate, Qdrant and Chroma built the vector storage. LangChain and LlamaIndex built the retrieval frameworks on top. Mem0 and Zep went specifically after agent memory. Sourcegraph spent years making large codebases searchable, and the agents themselves, Cursor, GitHub Copilot and Windsurf, each keep their own private notion of context.

All of that machinery points at documents and code, things that already existed before the agent turned up. The records of what agents actually did, and what failed, are produced in enormous quantity every day and thrown away.

The point: paid skill, owned skill, and now earned skill

The three-way split is worth holding onto, because it explains why this is genuinely new rather than knowledge management in a new box.

Paid skill is rented. You hire a senior engineer or a consultancy, you get their ability for as long as you pay, and it walks out with them. Every firm understands this and prices it.

Owned skill belongs to the firm: the processes, the collective judgement, the things written into documents, systems and culture. This is what people mean by competitive advantage, and companies have spent decades trying to hold more of it.

Earned skill is what builds up when the people you pay use AI on the company's real problems. It is specific to this firm and this job, it is produced constantly as a by-product of ordinary work, and until very recently it was not anything you could point at.

Earned skill is new not because nobody ever learned on the job before. It is new because, for the first time, the learning writes itself down automatically.

A person who spent six months finding out how a system really behaves kept that in their head, passed on a fraction of it in handover notes, and took the rest with them. An agent session produces a full record of the same discovery: what was tried, what failed, what the fix was, what finally worked. Nobody has to be talked into writing documentation. The record is already there.

And almost every company currently deletes it, or leaves it in a session history nobody can search, which amounts to the same thing.

Why this is an asset argument, not a savings argument

Glen's published numbers are efficiency numbers: about a fifth less time, about a third less cost. Those are real, and they are also the least interesting reason to buy it.

Nobody buys a company-wide system to save on AI costs. Those prices fall by themselves, and every finance director knows it. Sold as a way to cut costs, this is a nice-to-have competing against a price that is already falling on its own.

The stronger argument is that captured earned skill builds into something the company owns. Every task an agent finishes makes the next one cheaper, faster and more likely to be right, and that improvement stays with the firm rather than with whoever happened to be at the keyboard.

Be precise about what that means. Accounting rules will not let a company put this on its balance sheet, because things you build yourself generally cannot go there. But it behaves exactly like an asset. It builds up, it can be moved around inside the business, it survives people leaving, and anyone buying the company would look at it closely. Plenty of things that were worth money long before the accounts recognised them started out this way.

The education problem, and the tipping point

This will take a while to sell, and the reason is built into the problem rather than a failure of marketing.

Show a buyer a percentage saved and they compare it against other percentages saved, and treat the purchase as optional. A buyer who understands they are currently throwing away something they already paid to produce behaves completely differently, because the question stops being about efficiency and starts being about waste.

Getting a buyer from the first view to the second is the whole commercial job, and a demo will not do it. It probably takes a company looking back at six months of agent work and realising what it no longer has.

Once a few firms say that out loud, the argument should move fast. The moment this stops being about saving money and starts being about building something you own, it stops being a tool one engineering team picks up and becomes something the whole company installs, the way version control and document management eventually did.

Why it is bigger than it sounds

The quoted numbers are twenty-nine per cent less token spend and twenty-one per cent faster completion. The asset argument behind them is the larger one.

  • Agent transcripts are a genuinely new corpus. They did not exist two years ago, they are growing enormously, and almost nobody is keeping them.
  • Retrieving work is different from retrieving documents. The vector database wave solved the second one, which leaves this ground open rather than crowded.
  • This is the earned skill. The record of what an organisation has actually done becomes something it owns, alongside the skills it pays for and the ones it licenses.

What to watch

The number that matters is not cost saved. It is whether a task attempted a second time inside a company goes measurably better than the first, and whether that gap widens over months. That is what would prove earned skill is genuinely building up, rather than just being a bit easier to look up.

The second thing to watch is who signs the cheque. If it is an engineering manager, this is a developer tool with a good story. If it ever reaches whoever is responsible for what the company knows, then the asset argument has landed, and this is much bigger than it currently looks.