Research map / Multi-agent systems and games / LLM agents

Learning to Share: Selective Memory for Efficient Parallel Agentic Systems

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Learning to Share (LTS), a shared memory bank for parallel agent teams with an RL-trained Qwen3-0.6B controller that decides which intermediate steps to admit, cutting runtime while matching or improving accuracy on GAIA and AssistantBench.

Problem

Parallel agentic systems run multiple agent teams independently, so they repeatedly perform overlapping intermediate steps (searches, parsing, code), wasting computation and wall-clock time on long-horizon tool-use tasks.

Approach

A global shared memory of (summary, output) pairs accessible to all parallel teams, with a learned binary admission controller (LoRA-tuned Qwen3-0.6B) trained by stepwise policy gradient using group-relative advantages, usage-aware reward shaping, and a sparsity penalty; agents stay frozen.

Key results

• On GAIA with Qwen3-32B, LTS reaches 30.4% accuracy at 892s versus 24.8% and 1569s for M1-Parallel.
• On AssistantBench with GPT-5.1, LTS reaches 26.7% accuracy at 882s versus 24.0% and 1389s for M1-Parallel.
• Naively admitting all steps (LTS-AddAll) cuts runtime but lowers AssistantBench accuracy from 24.0 to 23.0.

Surprising findings

• Naive admit-all memory reduces runtime but hurts accuracy, so selective admission is necessary.
• Removing the sparsity loss leads to near-unconditional admission (98.9% saved fraction... i.e., nearly all steps admitted) with the longest runtime and no accuracy gain.

Limitations

• Memory is ephemeral and per-task, with no persistence across instances.
• Controller cannot delete or revise memory entries.
• Evaluated only on two benchmarks with 3 teams as the main setting.

Experimental setup

Datasets: AssistantBench GAIA

Benchmarks: AssistantBench GAIA

Metrics: accuracy wall-clock runtime memory recall cross-team recall memories saved

Baselines: MagenticOne M1-Parallel LTS-AddAll LTS-LLM

Builds on: M1-Parallel MagenticOne GRPO AutoGen

Closest papers by meaning

• G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems (NeurIPS 2025 · similarity 0.89)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.89)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.88)
• Learning Latency-Aware Orchestration for Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
• MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.88)
• Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.88)