The Pulse
MemComposer Improves Agent Results While Using Fewer Memory Tokens
MemComposer selects and orders retrieved memories for AI agents, improving task success in tests across three customer-service domains while reducing memory-token use. The study evaluates the method with two frozen agents and reports that i
AI.info Team · Written with AI ·
Similarity-based retrieval makes a straightforward bet: the memories closest to a new query are the ones most likely to help. A paper posted to arXiv on October 15 challenges that assumption, arguing that an agent’s results also depend on which memories it receives, how many it gets and the order in which it sees them. The researchers’ system, MemComposer, improved task success while reducing the number of memory tokens supplied to two tested agents.
The work comes from researchers at the University of Illinois Urbana-Champaign and Amazon. Rather than simply taking the highest-ranked search results, MemComposer assembles a query-specific sequence from retrieved memories, learning which combinations work through feedback from full task interactions.
Why the closest memory may not help
Many agent-memory systems store past experience, retrieve candidates for a new task and add those candidates to the model’s context. The authors say that similarity alone can miss useful guidance: a less similar experience may contain information that matters more for the task, while adding more memories does not guarantee better results.
“However, semantic similarity alone does not guarantee memory usefulness: downstream performance depends on how memories are selected and presented to the agent.”
The authors write in “MemComposer”
MemComposer treats memory selection as a joint decision. For each query, it chooses a variable number of candidates and sets their order; a fixed similarity-based retriever still supplies the candidate pool. The distinction matters: the system learns how to arrange retrieved material, not how to replace the underlying retrieval step.
Two frozen agents, three service domains
The team tested the method on τ²-bench, a benchmark of multi-turn customer-service tasks in retail, airline and telecom settings. Agents must follow domain policies, use tools and interact with a simulated customer. The experiments used two frozen executors, Qwen3-4B-Instruct and Ministral-3-8B-Instruct, and gave each method the same ten retrieved candidates.
With Qwen3-4B-Instruct, MemComposer reached an average task success rate of 50.8%, compared with 45.0% for the strongest baseline; its four-trial consistency score was 32.5%, versus 27.5%. With Ministral-3-8B-Instruct, it recorded a 63.5% average success rate, 4.6 percentage points above the strongest baseline, and a 40.0% consistency score. The method was best or tied for best in 10 of 12 domain-level comparisons.
Those gains came with lower memory-token use, not a demonstrated reduction in every kind of inference cost. Across the domains, MemComposer used 16.9% fewer memory tokens than the strongest baseline with Qwen3 and 4.9% fewer with Ministral. The paper reports that the pooled improvement in average success was statistically significant at the 95% level against every baseline; for the consistency metric, that finding did not hold against two baselines, Complexity and Memento.
Learning from completed interactions
The researchers train the composer with Experience Policy Optimization, or ExPO, a framework based on Group Relative Policy Optimization. It evaluates candidate memory sequences through complete interactions with a frozen agent, then uses task-success feedback to update the selection policy. A penalty for larger selections discourages unnecessary context, while another penalty covers invalid output formatting.
The system also showed transfer in the tested settings. A composer trained on retail tasks performed competitively when applied without further training to airline and telecom domains, and a composer trained with Qwen3 transferred to a Ministral executor. Those tests support the possibility that the learned selection strategy can travel beyond its original setup, but they do not establish performance across arbitrary agents or memory stores.
Evidence still limited to procedural memories
The study evaluates memories made from complete past task trajectories and reviews of those trajectories. It does not test whether the approach works as well with other forms of memory, such as stored facts, rules or a specific user’s history. The authors also say their two executors have 4 billion and 8 billion parameters, leaving open whether larger frontier-scale agents would benefit from a separate composer.
For now, MemComposer’s result is specific: on three simulated customer-service domains and two frozen agents, learning to select and order retrieved experiences beat the strongest tested alternatives while using fewer memory tokens. Whether the same trade-off holds for other memory types and larger agents remains untested.