SAGE: A Novelty Gate for Efficient Memory Evolution in Agentic LLMs

Researchers introduce SAGE, a memory management system for agentic LLMs that uses novelty detection to efficiently control when new facts are added, merged, or ignored. The approach reduces API costs and latency by 3.4× and 2.5× respectively while maintaining quality, addressing a critical gap in write-side memory control for long-context AI agents.

SAGE addresses a fundamental inefficiency in agentic LLM systems: the lack of principled write-side control for memory management. While prior research emphasized retrieval and storage mechanisms, this work tackles the equally important problem of deciding when and how to update memory. The system uses a von Mises-Fisher-based density estimator to score candidate facts against existing memory embeddings, categorizing them as clear additions, redundant operations, or uncertain cases requiring LLM judgment. This novelty-aware routing framework directly reduces computational overhead by avoiding unnecessary LLM calls during memory updates.

The research represents a maturation of long-context AI agent design. As LLMs are increasingly deployed in persistent reasoning loops, memory management becomes a bottleneck—each write operation carries API costs and latency penalties. SAGE's adaptive thresholding mechanism that tracks memory-store geometry addresses this by establishing domain-aware boundaries for novelty rather than applying static cutoffs. The empirical results demonstrate substantial practical benefits: 3.4× cost reduction and 2.5× latency improvement on GPT-4o-mini with minimal quality degradation, while skipping 16-18% of LLM calls across multiple backbones.

For AI developers and operators, this work signals that memory efficiency deserves engineering focus comparable to retrieval mechanisms. The drop-in compatibility of SAGE with existing memory systems like A-Mem and Mem0 enables rapid adoption. The technical approach—combining geometric density estimation with adaptive gating—provides a template for similar write-side optimization problems. Going forward, integrating novelty-aware write control into production agentic systems could substantially improve cost-performance tradeoffs as agent reasoning horizons expand.

• →SAGE reduces API costs by 3.4× and latency by 2.5× for memory write operations in agentic LLMs
• →Novelty detection via von Mises-Fisher density estimation enables routing decisions without unnecessary LLM calls
• →Adaptive thresholding based on memory geometry outperforms static novelty cutoffs across multiple models
• →The system skips 16-18% of LLM merge calls with minimal quality degradation on open-weight backbones
• →Drop-in compatibility with existing memory systems enables practical deployment in production agents