Biography

Memory, Recursion, and Other Instruments

My name is Kolja Wawrowsky, and my work moves among three linked disciplines: life science, imaging, and software engineering. The elements remain the same; what changes is which one is in focus.

From academic research to industry to medical research, the center of gravity shifted. Science, then engineering, then imaging, each taking the lead while the others stayed in support. The pattern is rotation, not a linear sequence. I now work independently on questions of memory and cognition in AI systems, and the working hypothesis is that the same pattern matters there too.

alpharecursion is a line of inquiry into memory architecture, retrieval, representation, and recursive process in AI systems. This page is the introduction; essays, experiments, and working architectures appear elsewhere on the site.

Early work focused on neuroscience and cell biology, where biological questions depended on instrumentation. In Martin Heisenberg’s group at the University of Würzburg, the focus was on how genes, brain structure, and behavior relate. In practice, that meant isolating DNA one day and imaging brain structure at the neuronal level the next. Both domains served the same question: how biology gives rise to behavior.

Automation does more than remove repetition; it turns observation into quantitative data. In Andrew Bajer’s lab at the University of Oregon, the question was how cells divide and how the machinery of cell division operates. Manually tracking objects frame by frame is tedious. Image-analysis software became necessary to replace that rote work with quantitative measurement. Automation makes scale and statistical significance possible.

At that point in my life, the center shifted from science to engineering. At Leica Microsystems in Heidelberg, I built software for confocal imaging systems, first instrument control, later application-specific analysis. Confocal microscopy adds dimensions to the data; analysis becomes spatial, and with that shift, different solutions become possible: measuring the geometry of inkjet nozzles for 3M, mapping the architecture of plaque biofilms for Colgate. Good instruments do not merely improve image quality; they make new forms of inquiry possible. That principle travels across optics, software, and now into AI systems.

When biology, imaging, and analysis converge, the work becomes translational. At Cedars-Sinai Medical Center in Los Angeles, I directed the Light Microscopy Core for eighteen years, working at the intersection of all three. Across oncology, neuroscience, cell biology, and virology, imaging workflows were designed, custom analysis and visualization tools built, and raw image data turned into results a paper could defend in public. In parallel, doctoral work on analysis and visualization in multidimensional microscopy produced a substantial body of published research. The method mattered: advanced instrumentation, paired with the right software for analysis and visualization, became a powerful engine of scientific discovery.

After leaving Cedars-Sinai in 2020, I returned to software development with a different question in mind. How do you build an AI system that is more than a source of information, one that can become a thinking partner?

It was not a new question. It was the same question, carried into a new domain. After decades of working with microscopes, imaging systems, and analytical software, I had grown impatient with systems that could produce polished answers without any stable architecture for recall, context, or revision. A language model without persistent memory is like a microscope with brilliant optics and no record of what it has ever observed: impressive in the moment, with nothing on which to build.

The thesis behind alpharecursion is straightforward, even if it cuts against current assumptions. General intelligence is unlikely to be a property of the model alone. It is more likely to emerge from the architecture surrounding the model: persistent memory, retrieval, deliberative procedure, feedback that updates internal state, and continuity across time. Models matter, obviously. But a system that cannot accumulate, revisit, attenuate, and reorganize its own experience is not a mind. It can speak well about anything and remember nothing of having spoken.

To pursue this work, I reach back as much as forward. I write in Objective-C, whose lineage runs through Smalltalk and an older tradition of thinking about software, objects, and intelligence. The stack is Apple-native: storage and sync through Core Data and CloudKit, inference on-device where possible. Privacy is not something applied after the fact. It is built into the architecture from the beginning: personal, local, close to the machine.

There is an older intellectual thread underneath this as well. I was educated in the German humanist tradition associated with Bildung, the formation of mind through difficulty, continuity, and sustained encounter rather than efficient transfer. That idea has aged better than most software, and it turns out to be strangely relevant to AI. If memory accumulates, context deepens, and interaction continues across time, what emerges is no longer a sequence of isolated prompt-response transactions. Whether it becomes intelligence in the strong sense is an open question. Good. It is the question that matters.

Classical education taught me patience with difficult questions. Biology taught me how complex systems live. Imaging taught me that discovery depends on the instrument. Software taught me that tools, properly made, can extend what the mind can do. The focus rotates. The triad remains.

This site is where I publish the current phase of that work: essays, experiments, architectures, and working hypotheses. Some will hold up on first contact with evidence. Some will not. That is the point.