Research map / Multi-agent systems and games / LLM agents
G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems
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 G-Memory, a three-tier graph memory (insight, query, interaction graphs) for LLM multi-agent systems that retrieves high-level insights and condensed collaboration trajectories, improving AutoGen, DyLAN and MacNet across five benchmarks.
Problem
Existing memory mechanisms for LLM-based multi-agent systems are overly simplistic: they ignore inter-agent collaboration trajectories and lack cross-trial, agent-specific customization, which hinders self-evolution.
Approach
A hierarchical graph memory over MAS executions: a query graph linked upward to a hyper-connected insight graph (generalizable lessons) and downward to an interaction graph (utterance logs). Retrieval does coarse embedding top-k plus 1-hop expansion, then LLM-based sparsification of relevant trajectories, and per-agent memory is initialized from these; memory is updated after each query.
Key results
- With Qwen-2.5-14B, G-Memory boosts MacNet on ALFWorld from 58.21% to 79.10% (+20.89%).
- With GPT-4o-mini and AutoGen, G-Memory averages 57.18% across five benchmarks vs 48.27% no-memory.
- Removing insights or interaction trajectories consistently lowers performance; interactions contribute slightly more (AutoGen avg drop 4.47% vs 3.95%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add G-Memory to AutoGen | AutoGen with GPT-4o-mini · PDDL | progress rate: +4.24 | Helped |
| Add G-Memory to MacNet | MacNet with Qwen-2.5-14B · ALFWorld | success rate: 58.21% to 79.10%; 20.89% | Helped |
| Add G-Memory to AutoGen | AutoGen with Qwen-2.5-14B · HotpotQA | exact match accuracy: 10.12% | Helped |
| Add role-specific high-level insights and fine-grained interaction trajectories (ablation) | G-Memory · AutoGen and DyLAN, average across benchmarks | average score: 4.47% and 3.82% | Helped |
| Add fine-grained interaction trajectories to insights-only memory (ablation) | G-Memory · AutoGen and DyLAN, average across benchmarks | average score: 3.95% and 3.39% | Helped |
| Increase query retrieval parameter to k=5 (ablation) | G-Memory on AutoGen · ALFWorld | success rate: 7.71%↓ | Hurt |
| Increase query retrieval parameter to k=5 (ablation) | G-Memory on DyLAN · FEVER | exact match accuracy: 2.5%↓ | Hurt |
| Use 2-hop expansion instead of 1-hop expansion (ablation) | G-Memory on AutoGen · PDDL | progress rate: 49.79% | Hurt |
| Use 1-hop query-graph expansion (ablation) | G-Memory on AutoGen · ALFWorld | success rate: 85.82% | Helped |
Surprising findings
- Most existing memory baselines (e.g., Voyager, MemoryBank) fail to consistently help MAS and can degrade performance on some tasks such as PDDL with AutoGen.
- Expanding retrieval beyond 1 hop on the query graph degrades performance.
Limitations
- Evaluated only on three domains and five benchmarks; no validation on e.g. medical QA.
- Memory invoked only at query onset in the implementation.
- Impact statement notes risk of amplifying incorrect reasoning if the underlying LLM is compromised.
Experimental setup
Datasets: ALFWorld SciWorld PDDL HotpotQA FEVER
Benchmarks: ALFWorld SciWorld PDDL HotpotQA FEVER
Metrics: success rate progress rate exact match accuracy token cost
Baselines: No-memory Voyager MemoryBank Generative Agents MetaGPT-M ChatDev-M MacNet-M
Builds on: AutoGen DyLAN MacNet Voyager MemoryBank Generative Agents
Closest papers by meaning
- MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.92)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.91)
- EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.91)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.91)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.