Researchers Propose Shared Selective Persistent Memory for Agentic LLM Systems

Researchers have proposed Shared Selective Persistent Memory (SSPM), a memory architecture for agentic large language model systems that enables agents to share and selectively retain knowledge across sessions. Agentic LLM systems that generate code through multi-turn tool use face a fundamental context problem: each session starts from scratch, forcing agents to re-derive solutions and relearn environment-specific patterns. SSPM addresses this by maintaining a persistent, structured memory store that agents can read from and write to across sessions, with a selective retention mechanism that filters noisy or outdated information. The system uses a consensus-based write protocol where multiple agents must agree before a memory entry is committed, and a decay-weighted retrieval that prioritizes recent, high-utility entries. In experiments on code generation benchmarks, SSPM reduced token usage by 18-32% and improved first-attempt success rates by 12-19% compared to stateless baselines, while the selective retention mechanism prevented memory bloat and maintained precision above 85% over extended deployment windows.