Shared Selective Memory Boosts Agentic LLM System Performance.

Agentic LLM systems that generate code through multi-turn interactions often struggle with a fundamental context problem: each new session effectively starts from scratch, losing valuable configuration choices, domain constraints, and tool-use patterns from previous interactions. Simply retaining full conversation histories is inefficient and can degrade performance by introducing irrelevant information. A new architecture, "shared selective persistent memory," addresses this by intelligently identifying and retaining four key categories of reusable context: task specifications, data schemas, tool configurations, and output constraints. Crucially, this memory is shareable across users with role-based access control, fostering collaborative reuse without redundant setup. Implemented in a collaborative workspace platform where LLM agents create and maintain git-versioned artifacts, this system achieved a 96% task completion rate, significantly outperforming systems without memory (79%) or with full history (71%). A complementary "zero-token data refresh" mechanism further reduces task time by 14x for recurring updates, and summary-driven generation cuts per-invocation token costs by 97x compared to raw data injection. Replication on public datasets confirmed generalizability, highlighting that naive full-history persistence can actively harm agent performance.