Notes Are Not Evidence: How Persistent Memory Creates Information Cascades in Llm Agents

Persistent memory lets language-model agents carry information across sessions, but it can create a feedback loop: an agent reads an earlier conclusion, treats it as evidence, and writes it back. This is an information cascade. In classical social learning, even Bayesian decision makers can ignore their own observations after seeing enough earlier decisions, locking into a belief that may be wrong. With persistent memory, successive sessions of one agent play those roles; a shared board extends the process across agents. We test this across 15 models and 299,937 API calls. In controlled binary tasks with stated observation reliability, appending each session's conclusion to memory makes Claude Opus 5 and Sonnet 5 persistently wrong in 20.0% of multi-session runs while confidence rises. Opus 5 discounts redundant conclusions yet still lets two matching stored conclusions outweigh a conflicting fresh observation. One assertively worded stored conclusion can even override its contrary observation. Replacing conclusions with a running tally of observations removes wrong lock-in for Opus 5 and raises final accuracy to 98.3%. On 12-agent Sonnet 5 boards, requiring an independent first judgment before reading others cuts wrong consensus from 19/80 to 3/80 boards (6.3-fold). Hiding the reliability disclosure substantially weakens note-following. Persistent memory can therefore make previous decisions behave like accumulating evidence, allowing additional sessions or agents to reinforce an error without adding independent information. Preserving evidence rather than verdicts and eliciting independent initial judgments substantially reduce this failure.