EGCForest: Structured Agent Memory Organization for Enterprise Group Chats
- Published
- Source
- arXiv
- Paper number
- 1346
- Field
- Agents / Memory
- arXiv ID
- 2610.19836
Key points
- The authors propose EGCForest, a framework that organizes enterprise group-chat logs into three complementary forests: Channel, Actor, and Artifact.
- The three forests share one immutable message log, while a Placement Map and sparse Bridges create cross-view connections and preserve traceability to original messages.
- Online retrieval runs in four stages of view-aware leaf recall, summary-guided supplementation, structural completion, and budget-aware assembly, requiring no LLM calls for retrieval.
- Across three enterprise-chat benchmarks, EGCForest consistently outperforms RAG and memory-system baselines, improving answer accuracy by up to 20.85 points over MemOS.
- On the internal Company data, the stage-wise contribution shows structural completion raising P@20 from 7.1 to 8.0 and accuracy from 47.9 to 53.3.
- Retrieval latency is 65 ms per query, the fastest among compared methods, with a 206 MB storage footprint.
Paper links
External research summaries. These are not HDATF publications or measured product results.