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

MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging

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 MemForest, a training-free agent memory compression framework that partitions memory into EventTrees via semantic-temporal clustering, progressively merges nodes along a maximum spanning tree with LLM-based fusion, and adds anchor-guided propagation retrieval.

Problem

Agent memory stores grow continuously, raising storage and retrieval costs; existing compression targets generation or token-level redundancy and does not exploit event structure and temporal continuity of textual/multimodal memories.

Approach

Partition memory nodes into event units using both global semantic similarity to cluster centers and local temporal continuity, model each unit as a maximum spanning tree, and progressively merge highest-weighted (degree-adjusted) edges via an LLM summarizer; retrieval propagates scores to temporal neighbors of anchor nodes.

Key results

• At 50% compression under Mem0, MemForest retains 97.1% of original performance (avg 59.3 vs 61.1) and with AGPR 98.3%, outperforming ToMe and pruning baselines.
• Under M3-Agent at 50% compression, MemForest retains 99.7% of performance on average; with AGPR it reaches 120.7%.
• Average retrieval speedup at 50% compression is 1.89x on Mem0 and 2.24x on M3-Agent.

Surprising findings

• With AGPR, M3-Agent performance exceeds the uncompressed baseline by about 20%, while Mem0 gains are small (0.2% without compression).
• Pruning methods such as StreamMeCo perform worse than random compression at over 50% compression on M3-Agent.

Limitations

• Merging relies on an external LLM (GPT-4o-mini), adding cost and dependence on a proprietary API.
• Gains from AGPR are framework-dependent and small on Mem0.
• Hyperparameters (alpha, beta, w, eta, gamma, L, lambda) are set by fixed values; sensitivity is only partly analyzed.

Experimental setup

Datasets: LoCoMo LongMemEval PersonaMem M3-Bench-robot M3-Bench-web

Benchmarks: LoCoMo LongMemEval PersonaMem M3-Bench-robot M3-Bench-web

Metrics: accuracy performance retention rate retrieval speedup B1 F1

Baselines: Random Pruning KMeans DART StreamMeCo Random Merging ToMe

Builds on: Mem0 M3-Agent ToMe DART StreamMeCo

Closest papers by meaning

• SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.91)
• AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.90)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.90)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.90)
• Mem-T: Densifying Rewards for Long-Horizon Memory Agents (ICML 2026 · similarity 0.89)
• AgentKVShift: Efficient KV Cache Reuse for Agentic Memory Systems (NeurIPS 2026 · similarity 0.89)