MemCon learns adaptive memory management for LLM agents

Researchers introduce MemCon, a framework that models memory operations as a Markov Decision Process to learn an online policy for adaptive retrieval and consolidation in LLM agents. Unlike static heuristics, this approach dynamically decides when and how much to retrieve based on context.

• MemCon wraps any existing memory implementation and learns from task-by-task binary feedback without pretraining or additional LLM calls.
• It uses a lightweight tabular contextual bandit with UCB exploration that converges within tens of tasks.
• Across 6 benchmarks, 3 agent frameworks, and 3 LLM backbones, MemCon outperforms baselines by up to 15.2 points in task success.
• The method reduces token consumption by 5--20% while improving performance.

MemCon addresses the bottleneck of fixed memory access by allowing agents to adaptively manage experience accumulation and plan reuse.