A simulation of synthetic consumers making daily purchase decisions, designed to study how LLM-based memory systems influence behavior over time.
What it does
Three synthetic consumers — Maya, Raj, and Elena — each with distinct personas, incomes, and personalities, browse a small retail catalog and make daily decisions via a shared BehaviorEngine backed by Claude Sonnet 4.6.
Every seven simulated days, a Reflection engine summarizes each consumer’s recent behavior into long-term memory. Those summaries feed into the next week’s decisions.
The surprising result
All three consumers converged on buying the same drip coffee for 80+ consecutive days, despite wildly different personas. The reflection loop had created an echo chamber: summaries described habits as identity, which made the next decision reinforce the habit, which made the next summary describe it as even more entrenched.
The fix
Reframing memory from third-person behavioral analysis to first-person diary entries broke the loop. Between-persona behavioral similarity dropped from 0.99 to 0.69. Cart abandonments increased 25×.
The format of memory mattered as much as the content inside it.
Write-up
Full experiment write-up: Why Is Everyone Buying the Same Coffee?