EGCForest: Structured Agent Memory Organization for Enterprise Group Chats

TL;DR

EGCForest introduces a hierarchical memory system for enterprise group chats that organizes messages across three interlinked forests: by channel, speaker, and artifact. It enables accurate retrieval of distributed evidence by combining message-level relevance with cross-view connections and hierarchical summaries. The framework achieves significant gains in accuracy over existing methods while supporting efficient real-time access.

Why It Matters

Enterprise agents often struggle to retrieve accurate, context-aware information from complex chat histories. EGCForest improves decision-making and task execution by reliably surfacing relevant evidence across threads, speakers, and deliverables. This enhances automation and knowledge management in real-world organizational workflows.

Key Findings

• EGCForest outperforms RAG and memory-based baselines on three enterprise chat benchmarks, with up to 20.85% higher accuracy than MemOS.
• The multi-view forest structure—Channel, Actor, and Artifact—captures temporal workflows, speaker authority, and artifact lifecycles effectively.
• Cross-view connections via a Placement Map and sparse bridges enable retrieval of evidence that spans multiple dimensions without losing context.