Shared Selective Persistent Memory for Agentic LLM Systems (research paper)

KEY POINTS

1. The authors propose "shared selective persistent memory," an architecture that preserves four reusable context categories (task specifications, data schemas, tool configurations, and output constraints) while discarding session-specific reasoning, and enables workspace-level sharing with role-based access control.
2. Implemented in a deployed collaborative workspace producing git-versioned artifacts, the approach yields higher task completion (96% vs 79% with no memory and 71% with full-history persistence), a 14× reduction in recurring-task time via a zero-token data refresh, and a 97× per-invocation token-cost reduction through summary-driven generation; replication on public datasets reported zero-token refresh success in 12/12 trials.
3. This addresses the fundamental context-loss problem for multi-turn, tool-using LLM agents by keeping compact, shareable context that improves task completion and efficiency while avoiding harmful stale-history bias.