From folder maps to living company memory

Every company we talk to has the same quiet problem. Their knowledge is real, but it is scattered: across Drive, Slack, inboxes, old documents, and people's heads. To a chat model, all of that is invisible. So the AI gives confident, generic answers that are wrong for them, and everyone slowly lowers their expectations.

memrelay exists to close that gap. But it did not start as a product. It started as a folder structure.

"We just need some structure"

The first version was almost embarrassingly simple: a folder-and-repo layout any company could drop its knowledge into and use through an AI assistant. Five tiers, from a single person up to an enterprise. Git as the source of truth. Markdown files a human could actually read.

That bare skeleton taught us the first lesson that still holds today:

The bottleneck was never the model. The bottleneck is that your AI has no context on your company.

Good structure helped. But structure alone is a filing cabinet. It does not remember, it does not connect, and it certainly does not answer questions.

The thing we kept rebuilding

A pattern repeated through the whole journey: we would introduce something, collide with reality, and rebuild it. The biggest recurring fight was deceptively boring: how does a non-technical person contribute knowledge without ever touching the machinery underneath?

We tried layers of automation. We tried policy files. We tried hooks. Each attempt failed in a new and instructive way, and each failure pushed us toward the same conclusion: the safety and the governance have to live on the server, not in good intentions. Writes should flow through review, deletes should never be destructive, and nobody should need to learn plumbing to add what they know.

From filing cabinet to memory

The turning point was reframing the whole thing. A knowledge base is a place you put things. Living memory is something a company can recall from, reason over, and watch grow. The same files, but now reachable by any AI client, structured enough to connect, and governed enough to trust.

That is the shape memrelay took: your knowledge stays yours, readable, hosted in the EU, and exposed to your AI through a single connection instead of a pile of copy-paste. You take the unglamorous step, structuring your knowledge and deciding what is true and who can see it, in minutes instead of months.

What we actually learned

• Structure is necessary but not sufficient. A perfect folder tree still cannot answer a question.
• Governance belongs on the server. If a guarantee depends on people remembering a rule, it is not a guarantee.
• Readable beats clever. Knowledge you cannot read is knowledge you cannot trust or correct.
• The step everyone skips is the whole job. Bringing knowledge in and deciding what is true is unglamorous, and that is exactly why AI keeps disappointing the companies that skip it.

If any of this sounds like your company, that is the gap memrelay was built to close. The models are already good enough. The missing piece was always your company's memory.