TL;DR: For about a year we have fed our own decisions, debates, and dead ends into a knowledge graph we call AIP. It now holds more than a hundred running threads of thinking, every claim traceable to its source, with disagreements kept rather than smoothed away. The payoff: new joiners can question the company directly, we stop relearning lessons we have already paid for, and the thinking stops disappearing when people do.
A few weeks ago someone new joined us and asked a simple question: why do we do it this way? I started to answer, got two sentences in, and realised the honest version was “because of a few conversations we had a year and a half ago, most of which I can only half remember.” For most of my working life that was the best anyone could do. It isn’t any more.
For about a year now we have been feeding our own work into a knowledge graph. Not the client work we get paid for, our own: the decisions, the arguments, the things we tried that didn’t work, the reasons behind the reasons. We call it AIP. It started as an experiment and it has quietly become the thing I check first.
Here is what it actually holds. More than a hundred ongoing “subjects”, which are just threads of thinking that gather over time. A subject isn’t a document or a meeting. It is a question we keep returning to, and underneath it sits every claim anyone has made about it, each one tagged with where it came from. Click on a claim and you land on the conversation or the file it was lifted from. Nothing floats free.
The part I didn’t expect to care about is that it keeps the disagreements. Most tools read your stuff and hand you back a tidy summary, which sounds helpful until you notice the summary has quietly deleted the bit that mattered: the point where two of us thought different things. AIP marks those. It will tell you a subject is “unsettled” when only one person has said something, “consensus” when separate sources agree, and it flags the open tensions instead of smoothing them over. It never tells me something is simply true. It tells me how well supported it is and leaves the deciding to me. That turns out to be a much more honest kind of help.
So what does this change, day to day?
Onboarding, for one. The new joiner who asked me “why do we do it this way” doesn’t really need me. They can ask the company directly and read the actual thread, contradictions and all, rather than a wiki page someone wrote once and never updated.
Then there is the thing I’d call not paying twice. Every team relearns its own lessons. You solve something, the person who solved it moves on or just forgets, and eighteen months later you solve it again from scratch. A graph that remembers is the cheapest insurance I know against doing that.
And brainstorms compound. We have a lot of good conversations that used to evaporate by the following week. Now the useful claims from them stick around and attach themselves to the right subject, so the next conversation starts further along instead of starting over.
I should be honest about the catch, because there is one. None of this works on its own. The graph is only as good as what we put into it, and the discipline of actually capturing the thinking, rather than just doing the work and moving on, is the hard part. We get it wrong plenty. But even a half-fed memory beats no memory, and the more we use it the more it pulls its own weight.
The bigger realisation is this. For years the unspoken rule in most companies has been that knowledge lives in people’s heads, and when people leave it leaves with them. We have flipped that to a slightly uncomfortable new rule: if it isn’t in the graph, it didn’t happen. That sounds severe written down. In practice it just means the thinking stops disappearing.
Which, if you have read anything else we have written lately, is the whole point. We keep saying the interesting move with AI isn’t doing yesterday’s work faster, it is going after problems you couldn’t touch before. “A company that actually remembers what it learns” was one of those problems for us. Turns out it was soluble. The reward, as ever, is that we have found some better ones.
That’s the update.