coglogic · bespoke cognitive systems
coglogic is a small studio for machine cognition: memory, context and the machinery of thought, made for the problem each one serves. We take on a few problems at a time and stay with each until it runs. The first of them is engram, a memory substrate for the people who design cognition.
What we make
The intelligence being built today rests on tooling made for a smaller problem. Memory forgets, context buckles, and the gap between what these systems could do and what their foundations allow keeps widening. We build the foundations: memory that keeps what it learns, context that stays whole under its own weight, and the machinery that composes the two into thought.
Nothing here comes off a shelf, and nothing is assembled from parts that almost fit. A system is specified before it is built, written down in full, and exercised hard once it runs. What we hand over is not a study or a proposal. It runs, against your data and inside your constraints.
Nothing off the shelf. Each system is shaped to the problem it was made to hold, and to the constraints around it: the data it may touch, the latency it must meet, the rules it has to live within. The shape of the problem is the shape of the system.
A small number of the right problems, given whole attention. Fewer engagements, longer arcs, harder problems seen through from the first sketch to the running system. We would rather finish three things well than start ten.
We begin where the reference implementations end. Much of what we build has none to lean on, so we write the reference: the definition first, in full, and then the code that is held to it. Working past the proven does not mean working loose.
The first system
A memory substrate that holds every form a system learns in. Made for the people who design cognition.
engram is defined in three parts and built as one: what it can hold and how it can be asked; how what it holds is kept, dated and recognised; and what it makes of what it is told. Each part is written in full, each is a lens on the other two, and nothing lands in the code that the three do not agree on.
What it does is on the next page: one memory across text, image, sound and signal; two clocks on every fact; confidence and provenance carried as values; recall that composes rather than fetches. What it is stays close to us. Not a concept, and not a claim. A system that runs.
A few problems at a time, each seen through.
If you are building cognition on tooling that was not made for it, write to us. We read every note ourselves, answer within days, and say plainly whether the problem is one we should take on.