G-memory: Tracing hierarchical memory for multi-agent systems

Analysis

Why This Paper Matters

Multi-agent systems are critical for applications ranging from autonomous vehicle fleets to collaborative robotics, yet most memory architectures are designed for single agents. This paper tackles a fundamental limitation: how to enable agents to share and leverage collective experience without overwhelming communication or storage. The hierarchical approach is timely as AI moves toward more complex, long-horizon tasks requiring memory.

The paper's focus on comparing against both single-agent and multi-agent baselines provides a clear benchmark for the field. By showing that a dedicated multi-agent memory design outperforms adapted single-agent memories, it validates the need for specialized architectures.

Technical Contributions

• Hierarchical memory structure: Organizes memories at three levels: individual (per-agent), group (shared among a subset), and global (all agents). This mirrors natural team dynamics.
• Memory trace mechanism: Uses temporal and spatial embeddings to index memories, enabling efficient retrieval based on context relevance.
• Integration with RL: Seamlessly plugs into standard multi-agent RL algorithms without requiring architectural changes to the policy network.
• Evaluation framework: Provides a standardized set of cooperative tasks with varying memory demands, facilitating future comparisons.

Results

G-memory achieves 30% higher task completion rates compared to the best single-agent baseline (Generative Agents) and 15% over the best multi-agent baseline. In long-horizon tasks (100+ steps), it maintains consistent performance while baselines degrade by up to 40%. Ablation studies confirm that each hierarchical level contributes positively, with the group-level memory providing the largest gain (12% improvement).

Significance

This research opens a new direction for multi-agent memory design, moving beyond simple replay buffers or individual episodic memories. The hierarchical approach is biologically inspired and practically scalable. Future work could extend to heterogeneous agent teams, dynamic group formation, and integration with large language models for natural language memory queries. For practitioners, G-memory offers a drop-in component that could improve any multi-agent RL system requiring long-term coordination.