Context Doesn't Travel

Last October I published a piece about a twenty-five-year-old idea. In 2001, Tim Berners-Lee and his colleagues described a web where information was not just written for people to read but structured so machines could understand what it meant. It was called the Semantic Web, and it did not happen.

I have been going back through what I published while we are building BeMo, checking each claim against what building has taught us. This one has a date on it now, because in September four software companies went and tested it in public.

What I claimed

I was clear about why the original idea failed. "For Berners-Lee's idea to work, every publisher and platform would have had to agree on a shared language for meaning." Nobody agreed. The dream stalled, and I wrote that it stalled for want of coordination rather than for want of a good idea.

Then I said AI had removed the need for the agreement:

"What made the Semantic Web impossible in 2001 was the need for humans to pre-tag everything with meaning. What makes Semantic AI possible in 2025 is that we no longer have to."

The machines could infer the structure themselves. No standards body, no shared vocabulary, no twenty years of waiting for the world to mark up its data. I ended the piece saying the future had arrived, "not on the open web, but inside the tools we use every day."

Read that last line again. I wrote the limitation into my own conclusion and did not notice it.

What September did

In one month, Salesforce, HubSpot and Glean each put the same word at the center of what they were building. The word was context. Blackbaud arrived at the same idea from another direction two weeks later. My co-founder, Becky Kern, wrote about that on Monday from the nonprofit's side, and about the people who have been describing the same problem for years without using anyone's word for it.

Here is the builder's half of it. Four companies now have a context. Each one is assembled from what sits inside that company's walls, and each one stays there.

That is not a complaint about any of them. It is the only thing a vendor can honestly build. A system can infer structure brilliantly from the material it can see, and no system can infer structure from material it cannot see. So each one can become very good at its own territory and stay silent about everything next to it.

Which means the agreement problem I declared solved was not solved. It moved.

In 2001 it was a problem between publishers. The whole web had to agree. In 2026 it is a problem between vendors, and the number of parties who would have to agree dropped from millions to four, or ten, or however many systems one organization happens to run. That is a much smaller problem. It is not a different one.

I had a confession to make somewhere in this piece, so here it is, and it is one sentence. I thought AI had dissolved the need for anyone to agree on meaning, and what it had actually done was shrink the agreement problem to the size of a vendor.

What that looks like at the other end

Nonprofits do not experience this as a problem of interoperability or data architecture. They experience it as having to hold the whole organization together themselves.

The practical form of that is the fifteenth login.

One executive director told us last year that she runs about fifteen tools and none of them talk to each other, and that she was in the middle of deduplicating four thousand records at the time. Give every one of those fifteen a context next year and she now has fifteen partial versions of her own organization, none of them aware of the other fourteen. The thing she wanted was one picture. What the market is building her is a better version of the problem she already has.

The part that should bother anyone building in this space is what happens when she switches. Much of that understanding was inferred rather than entered, so there may be no durable record of it for her to take with her. It was derived from her data, it lives as the vendor's work product, and when she leaves she takes her records and leaves the understanding behind. She rebuilds it somewhere else, from scratch, which is the exact thing all of this was supposed to end.

What changed in BeMo

I framed this in 2025 as a problem of agreement, and I would frame it differently now. Nobody needs Salesforce and HubSpot to agree on what a program is, and waiting for them to is the 2001 mistake with new logos on it. The thing that has to be true is narrower. The understanding an organization builds up about itself has to survive the boundary of whatever system inferred it.

We made a decision about that earlier than I could have explained it, and September is what made it explainable.

What BeMo keeps is the organization's own record: what it decided, what happened, and where each piece came from, with that provenance attached to the item rather than held in a model's head. The chat is grounded in that record and says so when it does not know. The record is the product. It is not a cache of inferences we happen to hold about you. Which means the organization can build on what it knows even when the person, the tool or the conversation changes.

The practical difference is what it means to leave. We say there is no lock-in by design, and the reason we can say it is that the thing worth taking is the record itself, which is yours and is written down. Making that record genuinely portable, so an organization can take its accumulated understanding with it, is what we are building now. I am not going to tell you it is done, because it is not, and this is a piece about a claim I made too early.

I would rather be held to the direction. If an organization's understanding of itself can only exist inside the tool that inferred it, then every switch costs the organization its memory, and a sector where switching costs you your memory is a sector where nobody switches and nobody improves.

The scorecard

Structured meaning did arrive through AI rather than through standards, roughly the way I said: conversations producing records instead of evaporating into text. That is the half I got right, and Edition 4 already scored it. What did not hold is the sentence I was proudest of, that we no longer have to agree. True inside one product. False across any two.

The delta, in one sentence. I wrote that AI had removed the need for anyone to agree on what things mean, and building taught me it only moved the agreement problem from the whole web down to the boundary of a single vendor, which is why what BeMo keeps is a written record the organization owns rather than an understanding only we can hold.

Then: the dream arrived, inside the tools we use every day. Now: inside is the problem.