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

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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 MRAgent, an LLM agent with a Cue-Tag-Content associative memory graph and active, LLM-guided multi-step memory reconstruction, evaluated on LoCoMo and LongMemEval with gains over memory baselines and lower token cost.

Problem

LLM agents struggle to reason over long interaction histories because existing memory systems use static retrieve-then-reason pipelines (top-k similarity or fixed N-hop graph expansion) that cannot adapt to evidence discovered during inference.

Approach

Represent memory as a heterogeneous Cue–Tag–Content graph where tags mediate cue-to-content links, and let the LLM iteratively select traversal actions (forward and reverse) and prune branches based on accumulated evidence, framed as active memory reconstruction with a theoretical expressivity separation from passive retrieval.

Key results

• On LoCoMo with Gemini backbone, overall J rises from 68.31 (Mem0, best baseline) to 84.21 with MRAgent (about 23% relative gain).
• On LoCoMo with Claude backbone, overall J is 88.32 for MRAgent versus 78.61 for LangMem (best baseline).
• On LongMemEval with Gemini, overall J is 72.95 vs 54.92 for MemoryOS; token use is 118k per sample versus 632k for A-Mem and 245k for Mem0, with runtime 586s versus 533s for Mem0.

Surprising findings

• Increasing the parallel retrieval budget cannot substitute for deeper reconstruction depth, per the authors' appendix experiments.
• On LoCoMo multi-hop queries, the paper's own ablation shows reasoning-enabled variants gain most, while removing semantic memory degrades performance.

Limitations

• Reconstruction cost grows with traversal depth, so multi-step queries incur higher latency than single-shot retrieval.
• Static memory construction does not update or consolidate memory, so the graph grows monotonically with storage overhead.
• Theoretical separation relies on a constructed binary-tree needle task rather than natural language memory tasks.

Experimental setup

Datasets: LoCoMo LongMemEval

Benchmarks: LoCoMo LongMemEval

Metrics: F1 LLM-Judge accuracy (J) evidence recall token consumption runtime

Baselines: RAG A-Mem MemoryOS LangMem Mem0

Builds on: A-Mem Mem0 MemoryOS LangMem Zep

Closest papers by meaning

• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.91)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.91)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.91)
• EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective (NeurIPS 2026 · similarity 0.90)
• E-mem: Multi-Agent Based Episodic Context Reconstruction for LLM Agent Memory (ICML 2026 · similarity 0.90)
• Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (NeurIPS 2026 · similarity 0.90)