Salesforce study: Memory compounds mistakes in self-improving agents

rohanpaul_ai · x · 2026-08-22

Salesforce research highlights the fragility of memory-based self-improving agents. While ReasoningBank improved performance with the default (easy-to-hard) task order in WebArena, performance dropped when tasks were shuffled. This is because agents compound bad lessons (e.g., recommending impossible APIs) alongside good ones, leading to instability in 71% of cases.

Original post →

More from coding & agent

coding & agent channel →