A retrospective note. This essay predates the research lab in its current form; it records the working observations that, over time, turned into a research question.
The recurring failure was never the math
For close to two decades, the work was building systems that decide: rule engines, statistical models, real-time pipelines, process maps for organisations in the middle of transformation. Different industries, different stacks, the same lesson repeated often enough that it stopped looking like coincidence.
The expensive failures were almost never failures of computation. The model was usually fine. What broke was continuity. A system made a good decision in March and an inconsistent one in September, because nothing connected the two. A team solved a problem, reorganised, and solved the same problem again eighteen months later from scratch. The knowledge existed — it simply had nowhere durable to live.
We had, in other words, a memory problem dressed up as a dozen other problems.
Storage is not memory
The instinctive fix is to store more. Log everything, index it, make it searchable. Most organisations we worked with had already done this, and were no better off for it. A full archive that no one can interpret is not memory; it is sediment.
Human memory does something that storage does not. It decides what matters. It strengthens what is used and lets the rest fade. It links an event to its context — when it happened, what surrounded it, what it resembled. It reconsolidates: each time a memory is recalled, it can be revised in light of what is now known. None of these are retrieval operations. They are operations on the structure of what is remembered.
A system that only stores treats every record as equally important forever. That is not a neutral choice. It is the choice that guarantees the signal will eventually drown.
What the practitioner's view misses, and what it sees
We make no claim here to neuroscience. The observation was practical and came from the outside: the systems that aged well were the ones that had some discipline about forgetting and connecting, even if it was implemented by hand, in a wiki convention or a review ritual rather than in code.
That is the uncomfortable part. The good behaviour was always manual, always fragile, always the first thing to go when a team got busy. No one had built it into the substrate. The question that kept returning — and that we did not yet have the means to pursue — was whether the structure of memory could be made part of the system itself, rather than a habit layered on top of it.
We did not have an answer in 2022. We had a sharpened question, and the growing suspicion that the place to look for the answer was not in software engineering at all, but in how biological memory is organised. That suspicion is where the later work begins.
Next in this series: Why Software Forgets — and Why Retrieval Is Not Memory.