Memory Is a Configuration Problem: 5,474 Days Inside an AI Agent's Synthetic Lifetime

A pre-registered study of a simple question: if you give an AI agent a memory, and let it write that memory itself one day at a time, does it slowly become more human? I built Brandy, a billing-support agent, and had her keep a daily journal across a compressed fifteen-year synthetic lifetime (about 5,500 entries), accumulating self-authored memory through a standard extraction pipeline, at four memory depths (0, 5, 10, and 15 years). The prediction that memory would make her more empathetic was falsified. Instead, memory destabilized her. She fabricated more, grew no more consistent, and her self-narrative slowly came apart in specific, measurable ways: a pet that changes species, a move that never happens, a world that never ages across fourteen calendar years. Every failure traces back to one mechanism. The memory kept her work and dropped the rest of her life. The paper argues that memory is not a feature you switch on but a configuration problem, and that the thing you actually have to configure is relevance: what gets kept and what gets dropped. This project holds the full paper, the 5,474 generated journals, the analysis code, the pre-registered predictions, the blind-judging records, and the scripts that produced every figure. It was built and analyzed in collaboration with Claude (Anthropic's Opus 4.8, 1M-token context).