Current LLM agents do not retain task history, limiting their ability to learn from past interactions. Existing memory systems each use a single representation format like trajectories or skills, but no single format works best across all tasks. The authors propose MemAgent, which treats memory as a routing problem: a memory agent decides which of several heterogeneous memory providers to query for retrieval, whether to inject short-term memory during execution, and which providers should store the experience. Their architecture includes content-aware probing before retrieval, a short-term memory gate, and selective multi-provider storage. A training pipeline generates phase-specific supervision for routing decisions. In experiments across three benchmarks (GAIA, WebWalkerQA, xBench-DS), MemAgent achieves an average accuracy improvement of 10.0% over any individual memory method, with less than 0.3% routing overhead and a 12% reduction in average task steps.
Routing Across Memory Types Makes LLM Agents More Accurate and Efficient
MemAgent learns when to retrieve from, write to, and inject context from multiple memory representations.
Chinese Tech
Yongxian Wei · Yilin Zhao · Runxi Cheng · Xinrui Chen · Chun Yuan · Yaoru Wang · +2 more
Tsinghua University · Tencent
Research Digest··2 min read
In one line
A routing agent that selects among heterogeneous memory providers improves LLM agent accuracy by 10%.
Before you read this
Three ideas this paper builds onagent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.241 papers →model routingDirecting each query or subtask to the most suitable model to optimize cost, accuracy, or quality.89 papers →memory consolidationSummarizing or structuring past information to reduce token usage while preserving task accuracy.40 papers →
Read it as
Wei et al. tested 13 agent-memory methods and found that no single representation performed consistently across three benchmarks.
The authors evaluated 13 memory methods spanning representations such as prior trajectories, reflections, skills and structured knowledge on GAIA, WebWalkerQA and xBench-DS.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.