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

MemCoRe: Recovering Evidence from Progressively Compressed Factual Knowledge for Agent Memory

NeurIPS 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 MemCoRe, an agent memory system that organizes memory as a compression hierarchy (notes, keywords, topic groups) with gated merge/link/append construction and tri-pathway retrieval plus iterative evidence refinement, evaluated on LoCoMo with GPT-4o-mini, GPT-4o, Qwen3-8B and Qwen3-14B.

Problem

LLM agent memory must compress growing interaction histories while keeping relevant evidence retrievable; existing memory systems either retain redundancy that buries evidence or compress too aggressively and lose it.

Approach

Organize memory as a multi-level compression hierarchy (evidence notes, entity keywords, topic groups) with LLM-gated merge/link/append consolidation, and retrieve across levels via topic, keyword, and associative-neighbor channels, framed through an adapted information bottleneck objective.

Key results

• On LoCoMo with Qwen3-8B, MemCoRe reaches 38.62% average F1, 3.53 points above the second-best LightMem.
• With GPT-4o-mini, MemCoRe attains 43.76% average F1, 1.79 points above A-MEM.
• With GPT-4o, MemCoRe reaches 44.39% average F1, only 0.27 points above LoCoMo baseline, which the authors note may not be statistically significant.

Surprising findings

• On Qwen3-14B Adversarial, the LoCoMo baseline (49.17 F1) substantially outperforms MemCoRe (26.59 F1), which the authors attribute to lower-compressed context aiding distractor rejection.
• Removing link operations hurts more than removing merge operations (33.57 vs 34.79 F1), suggesting relational structure matters more than redundancy reduction.

Limitations

• Construction is slow due to LLM-based semantic evaluation at each step.
• Evaluated only on LoCoMo; no other benchmarks or multimodal settings.
• Information bottleneck is an analytical framework rather than directly optimized.

Experimental setup

Datasets: LoCoMo

Benchmarks: LoCoMo

Metrics: F1 BLEU-1 Recall Hit Rate

Baselines: LoCoMo ReadAgent MemoryBank MemGPT A-MEM Mem0 LightMem

Builds on: A-MEM Leiden algorithm Reciprocal Rank Fusion information bottleneck

Closest papers by meaning

• Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation (NeurIPS 2026 · similarity 0.89)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.88)
• MEME: Multi-Entity & Evolving Memory Evaluation (NeurIPS 2026 · similarity 0.87)
• MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging (NeurIPS 2026 · similarity 0.87)
• SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.86)
• MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.86)