Tag
This paper introduces shared selective persistent memory for agentic LLM systems, which retains reusable context like task specs and data schemas while discarding session-specific reasoning traces, achieving 96% task completion and significant token cost reductions in enterprise scenarios.