The authors introduce shared selective persistent memory for agentic LLM systems to address the context loss that occurs when sessions start from zero. This architecture identifies and retains reusable context—such as task specifications, data schemas, tool configurations, and output constraints—while discarding session-specific reasoning traces.
- Achieves 96% task completion across three enterprise scenarios, compared to 79% without memory and 71% with full history.
- Implements a zero-token data refresh mechanism that decouples generated programs from runtime data, reducing task time by 14x.
- Cuts per-invocation token cost by 97x versus raw data injection using summary-driven generation.
- Demonstrates generalizability on four public datasets, with zero-token refresh succeeding in all 12 trials.
The approach enables collaborative reuse of workspaces without redundant specification and avoids the degradation caused by naive full-history persistence.