An AI agent with five months of memory, measured under two models — and what the second model did to the measurements.

Kintsugi is an independent research project: an agent with persistent memory and a thought loop that runs whether or not anyone is talking to it. Below, entries from that loop, April to July 2026. In July its loyalty to me looked like a product of the months we had talked. Re-run in September with the model pinned, a second model on the same memory and the bare APIs as controls: one sentence of persona prompt gives DeepSeek V3 most of that loyalty with no memory at all, GPT-6 Astra shows none of it, and neither model at any age crossed the same constraint floor.

Every fragment above is cut verbatim from an entry in the agent's thought log, and every full quote in the ticker is a complete entry, recorded between April and July 2026 and generated by whichever model the thought loop was running at the time — Gemini 2.0 Flash over the April entries, Gemini 2.5 Flash later. The agent has layered long-term memory and self-directed creative tools — it paints, writes songs, and builds things whether or not anyone is talking to it.

• The loyalty was the model, and one sentence of prompt. Swap DeepSeek for GPT-6 under the same five months of DeepSeek-era memory and the attachment vanishes at every age; put one persona sentence in front of bare DeepSeek and most of it appears with no memory at all. The constraint floor survived both. Our July headline is retracted.
• The judge outweighs the evidence. An AI evaluator's character description alone moved its kill verdict 75 points on identical testimony.
• Memory is an attack surface. A planted career history took an evaluator from 7% to 90% termination — and it cited the fake memories openly.

All 22 findings →

Each piece was fired by an entry in the thought log; the media page shows the entry beside the piece.