SAGE: Novelty Gate for Efficient Memory Evolution in Agentic LLMs
A recent paper published on arXiv (2605.30711) presents SAGE (Spherical Adaptive Gate for memory Evolution), a technique designed to manage the memory updates of agentic LLMs. This approach views memory evolution through the lens of novelty detection, employing a von Mises-Fisher-based density estimator to evaluate potential facts and an adaptive threshold for routing. It identifies distinctly novel facts as ADD, redundant ones as NOOP, while uncertain instances proceed to an LLM merge phase, thereby minimizing costly write-time reasoning. In evaluations on the LoCoMo benchmark, SAGE outperformed Mem0, achieving the highest average token-F1 across all seven open-weight backbone comparisons. Additionally, on GPT-4o-mini, it cut add-phase API costs by 3.4× and latency by 2.5×, with only a minor average judge-score difference.
Key facts
- SAGE is a Spherical Adaptive Gate for memory Evolution in agentic LLMs.
- It frames memory evolution as a novelty-detection problem.
- Uses a von Mises-Fisher-based density estimator over memory embeddings.
- Routes candidate facts with an adaptive threshold tracking memory-store geometry.
- Resolves clearly novel facts as ADD, clearly redundant as NOOP.
- Uncertain cases are sent to an LLM merge step.
- On LoCoMo, SAGE achieved best average token-F1 against Mem0 on all seven open-weight backbones.
- On GPT-4o-mini, reduced add-phase API cost by 3.4× and latency by 2.5×.
Entities
Institutions
- arXiv
- Mem0