Shared Selective Persistent Memory for Agentic LLM Systems

2026-07-13 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Software Development & Engineering · Depth: Expert, extended

Summary

Apple Inc. researchers introduce "shared selective persistent memory," a novel architecture for agentic LLM systems designed to overcome the stateless nature of current multi-turn tool use. This system selectively retains four categories of reusable context—task specifications, data schemas, tool configurations, and output constraints—while actively discarding irrelevant session-specific reasoning traces. Implemented in a deployed collaborative workspace platform, it enables git-versioned artifact management and a zero-token data refresh mechanism. Across three enterprise deployment scenarios, this memory architecture achieved a 96% task completion rate, significantly higher than 79% without memory and 71% with full history persistence. The zero-token data refresh mechanism reduced task time by 14x for recurring data updates, and summary-driven generation decreased per-invocation token cost by 97x compared to raw data injection. A replication on four public datasets confirmed the generalizability of these findings.

Key takeaway

For AI Architects designing agentic LLM systems for enterprise workflows, prioritize selective memory architectures over full history persistence. Implementing shared selective persistent memory, which retains task specifications, data schemas, tool configurations, and output constraints, significantly boosts task completion rates to 96% and enables 14x faster recurring data updates. Avoid naive full-history approaches, as they actively degrade performance and bias agents with stale reasoning traces.

Key insights

Selective persistent memory for agentic LLMs improves performance by retaining reusable context and discarding session-specific reasoning.

Principles

• Persist "what" (declarative context), discard "how" (procedural traces).
• Irrelevant context actively degrades LLM generation quality.
• Decompose reusable knowledge into compact, stable categories.

Method

The system composes a session prompt from four persistent memory categories: task specifications, data schemas, tool configurations, and output constraints. Data schemas are injected into the user message.

In practice

• Implement zero-token data refresh for recurring tasks.
• Use statistical data profiles instead of raw data injection.
• Version artifacts with Git for risk-free exploration.

Topics

• Agentic LLM Systems
• Persistent Memory
• Context Management
• Code Generation
• Zero-Token Data Refresh
• Collaborative Workspaces
• Enterprise AI

Code references

• langchain-ai/langchain
• crewAIInc/crewAI
• features/copilot

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Shared Selective Persistent Memory for Agentic LLM Systems — Selective, shared, persistent memory significantly boosts LLM agent performance and efficiency by retaining only relevant context and enabling reuse.
• A Practical Guide to Memory for Autonomous LLM Agents — Memory architecture, particularly its management, is more critical to LLM agent performance than model choice.
• Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents — Dual-trace memory encoding enhances LLM agent recall by adding contextual "scene traces" to factual memories.
• Useful Memories Become Faulty When Continuously Updated by LLMs — LLM-based memory consolidation can degrade utility, making agents forget previously learned solutions.
• How to Optimize Your AI Coding Agent Context — Effective context management is paramount for optimizing AI coding agent performance and engineer efficiency.
• Engineering Memory for AI Agents: A Practical Beginner’s Guide — AI agent memory selectively retains past interaction details to enable contextual, multi-step task execution.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.