An ongoing personal experiment

What should an AI remember?

Impressionist is my small, visible way of exploring that question. A language model paints a photograph through tools, with limited attention, a finite token allowance, and memories that don’t last.

Look. Make a mark. Look again.

The model chooses where to look on the reference or canvas, which brush to use, its color and pressure, and the path of each stroke. MyPaint renders those decisions onto one persistent surface. It can overpaint or erase to revise its work.

Reference study and painting are separate model calls. A reference-visible call has no stroke tool: it can only write memory and choose a next view. Stroke-generating calls never contain a fresh reference image, only retained fading memories, written state, and optional canvas feedback. Older archived runs predate this rule and may have chosen strokes from a fresh glimpse. It sees one selected observation at a time. Looking away can leave a small visual memory in one of three slots. Those memories progressively lose spatial detail and color precision, then expire. A short notebook and recent intentions provide another kind of continuity.

A painting with a budget

The current series uses Claude Sonnet 5.5 with a 120,000-token allowance per painting, counting repeated inputs, images, reasoning, and outputs. That makes observation and written memory part of the artistic tradeoff: looking again costs resources too.

Each canvas has a daily cutoff at 6 p.m. Eastern, when the next photograph begins. A model that finishes early, or reaches its resource limit, leaves its painting on the easel until that handoff. The five photographs in this series are my own.

What is real today—and what comes next

This is a working tool-use and memory experiment, not a model I have trained. The fading images are produced by a programmed degradation process; they are not diffusion-generated memories. The model does choose its observations, notes, and marks. The harness sets the tools, limits, and pace.

The research direction is to train useful memory writing: can a model learn what to preserve for its future self, and can a visual memory serve a different purpose from a text summary? Coding agents use notes and context compaction to keep working over long tasks. Painting makes the same questions tangible.

The history preserves actual canvas snapshots and, where recorded, the observations and notes available to the model. These are visible working notes, not private chain-of-thought. Older studies may have only a finished image; missing frames are never invented.

Built by Isaak Hernandez. An evolving project, with experiments and imperfect paintings left visible.