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.
More from coding & agent
- FlowEvo: Self-Evolving Agents through Co-Evolution of Workflows and Skills — SoutheastU · 2026-08-22
- AI Agents Fully Automated Bug Fix and Code Merge — markjeffrey · 2026-08-22
- ClovaLink: Open-source Rust enterprise file management system — tom_doerr · 2026-08-22
- Open Source 'Computer History' Lets Agents Remember Paths via Accessibility Trees — altryne · 2026-08-22
- Claude Code now supports starting remote sessions from mobile — majidmanzarpour · 2026-08-22
- Crafting 1,000-line Bash installers quickly with a Claude Code Skill — doodlestein · 2026-08-22