PaperlayerView exact revision on arXiv

Source-linked AI summary

EGCForest: Structured Agent Memory Organization for Enterprise Group Chats

Lingxi Cui, Jinhao Ruan, Wenqing Wei, Guo Yi, Li Rao, Ke Chen, Lidan Shou

arXiv:2610.19836v1cs.CL

TL;DR

Enterprise-chat evidence is difficult to retrieve reliably because it spans interleaved threads, channels, and time, with meaning shaped by the speaker. EGCForest organizes an immutable message log into three connected forests and uses staged retrieval to assemble evidence. Across three benchmarks, it consistently outperforms RAG and memory-system baselines, with gains of up to 20.85 points over MemOS.

  • Problem

    Enterprise-chat evidence is distributed across threads, channels, and time, while speaker context shapes meaning and authority, making reliable retrieval challenging.

    Source: §1 · p. 1

  • Method

    EGCForest organizes an immutable message log into Channel, Actor, and Artifact forests and retrieves evidence through summary-guided and cross-view stages.

    Source: §1 · p. 2, §3 · p. 3, §3.2 · p. 5

  • Results

    Across three benchmarks, EGCForest consistently improves over strong RAG and memory-system baselines, outperforming MemOS by up to 20.85 points.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Experiments on three benchmarks show consistent improvements across diverse question categories for traceable organizational evidence organized across time, participants, and artifacts.

    Source: §5 · p. 8

    Evaluation uses question answering on three enterprise-chat benchmarks, leaving benefits for broader agent workflows such as task coordination and long-term planning to be explored.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Enterprise group chats record project progress, decisions, task ownership, and other organizational information valuable to enterprise agents. Reliable retrieval is challenging because evidence spans interleaved threads, channels, and time, with meaning and authority conditioned on the speaker. Neither RAG's independent chunk retrieval nor memory systems' summary-based consolidation can reliably recover such distributed evidence. We present EGCForest, a novel multi-view hierarchical memory framework for enterprise group chats. EGCForest organizes an immutable message log into three complementary forests: a Channel forest for temporal workflows, an Actor forest for speaker-conditioned context, and an Artifact forest for deliverable lifecycles. Cross-view connections are enabled by a Placement Map and sparse structural bridges. Built on this memory backbone, EGCForest retrieves and assembles evidence by jointly leveraging message-level relevance, hierarchical context from forest summaries, and cross-view links, recovering distributed evidence. Experiments on three enterprise-chat benchmarks show that EGCForest consistently outperforms RAG and memory-system baselines, improving answer accuracy by up to 20.85 points over MemOS while supporting efficient online retrieval.

1 Introduction

Enterprise group chats contain valuable organizational knowledge, but extracting it as reliable, traceable agent memory remains open because evidence is distributed across threads, channels, and time and depends on speaker context. EGCForest organizes that evidence in three connected forests and reports improvements over RAG and memory-system baselines across three benchmarks.

  • Enterprise group chats contain project progress, ownership assignments, and design rationales, but extracting this knowledge with traceable evidence remains an open problem.Source: §1 · p. 1
  • RAG may miss distributed evidence by retrieving independent chunks, while summary-based memory can obscure or conflate evidence from distinct threads and decisions.Source: §1 · pp. 1–2
  • Participant roles and responsibilities shape message meaning and authority, so speaker identity alone as metadata may miss evidential relevance.Source: §1 · p. 2
  • EGCForest organizes a shared message log into Channel, Actor, and Artifact forests, linked by a Placement Map and sparse Bridges for cross-view evidence retrieval.Source: §1 · p. 2
  • Across three benchmarks, EGCForest consistently improves over strong RAG and memory-system baselines, outperforming MemOS by up to 20.85 points.Source: §1 · p. 2

2 Preliminaries

The paper formulates enterprise-chat memory as retrieving compact evidence for a user’s query, and situates this task against memory and retrieval work that does not fully use group-chat structure.

2.1 Problem Formulation

  • A chat history is an ordered, immutable sequence of messages, and the task is to retrieve compact evidence for a natural-language query posed by an asking user.Source: §2.1 · p. 2

2.2 Related Work

  • Prior memory systems often compress single-user conversations, while retrieval and conversation-understanding work generally does not fully exploit group chats’ native reply, role, and cross-channel signals.Source: §2.2 · pp. 2–3

3 Methodology

EGCForest structures an immutable enterprise-chat log in complementary views and retrieves query evidence through a staged process that combines message relevance, summaries, cross-view links, and budget-aware assembly.

3.1.1 Three-View Forest Architecture

  • EGCForest builds Channel, Actor, and Artifact forests over an immutable message log, organizing temporal workflows, participant context, and deliverable lifecycles.Source: §3 · p. 3, §3.1.1 · pp. 3–4

3.1.2 Cross-View Connections

  • A Placement Map links a message’s placements across views, while sparse reply and identifier Bridges connect distinct messages for cross-view navigation.Source: §3.1.2 · pp. 4–5

3.2 Forest Retrieval

  • Retrieval produces compact evidence under a fixed token budget through four stages: view-aware leaf recall, summary supplementation, cross-view completion, and context assembly.Source: §3.2 · p. 5

3.2.1 S1 View-aware Leaf Retrieval

  • S1 fuses BM25 and dense message rankings, projects anchor messages to forest placements, and ranks trees using complementary view-aware signals.Source: §3.2.1 · p. 5

3.2.2 S2 Summary-guided Supplementation

  • S2 retrieves relevant internal-node summaries, expands selected nodes into bounded descendant leaves, and adds ancestor summaries as navigation headers.Source: §3.2.2 · p. 5

3.2.3 S3 Structural Completion

  • S3 completes evidence through bounded Placement Map joins for queries with at least two resolvable entities and limited reply or identifier bridge hops.Source: §3.2.3 · pp. 5–6

3.2.4 S4 Budget-aware Assembly

  • S4 deduplicates candidates while preserving stage order and fixed allocations, stops at the token budget, and renders selected messages verbatim with lightweight metadata.Source: §3.2.4 · p. 6

4 Experiments

Across three benchmarks, EGCForest achieves the highest average accuracy, while component and retrieval studies examine its views and retrieval stages. The experiments also report retrieval quality and efficiency.

4.1 Experimental Setup

  • Experiments use EverMemBench, GroupMemBench, and Company, with answer accuracy reported for all and evidence Precision@k and Recall@k additionally reported for Company.The setup lists RAG and agent-memory baselines and specifies GPT-5 for GroupMemBench and GPT-4.1-mini for the other two benchmarks.Source: §4.1 · p. 6

4.2 Q1: Main Results

  • EGCForest exceeds the strongest baseline in average accuracy by 4.77, 8.59, and 12.61 percentage points across the three benchmarks.The benchmarks are EverMemBench, GroupMemBench, and Company.Source: §4.2 · p. 6, Table 1 · p. 7

4.3 Q2: Component Study for EGCForest

  • The full model’s weighted average is best on both datasets; adding Actor improves Company Fine-grained Recall by 8.08 points and Memory Awareness by 5.32 points.Adding Artifact improves Profile Understanding by 6.61/2.88 points on GroupMemBench/Company; adding it to Channel+Actor improves weighted average by 0.93/3.37 points on GroupMemBench/Company.Source: §4.3 · pp. 7–8, Figure 3 · p. 8

4.4.1 Comparison with Standard Retrieval

  • EGCForest has the highest Precision@k at every retrieval budget while maintaining competitive Recall@k; GraphRAG has the highest Recall@20 but lower precision.The figure reports retrieval performance on Company.Source: Figure 4 · p. 8, §4.4.1 · p. 8

4.4.2 Stage-wise Contribution

  • At the same context budget, retrieval stages progressively add evidence, with S4 raising answer accuracy from 47.9 to 53.3.S1 reaches 40.3 R@20; S2 raises R@20 from 40.3 to 40.8 and accuracy from 45.4 to 45.8; S3 raises P@20 from 7.1 to 8.0 and R@20 from 40.8 to 44.5 through Placement Map joins and bridge hops.Source: Table 2 · p. 8, §4.4.2 · p. 8

4.5 Q4: Efficiency

  • EGCForest’s three-forest construction costs more offline than lexical and dense indexing but less than GraphRAG, MemoBase, and Mem0; online retrieval takes 65 ms per query.Storage is 206 MB; answer usage averages 2,365 input and 1,300 output tokens per query.Source: Table 3 · p. 8, §4.5 · p. 8

5 Conclusion

EGCForest organizes immutable message evidence through complementary Channel, Actor, and Artifact forests, and experiments show consistent improvements across diverse question categories. The evaluation is limited to question answering on three benchmarks; broader agent workflows and multimodal inputs remain future directions.

5 Conclusion

  • Experiments across three benchmarks demonstrate consistent improvements across diverse question categories.Source: §5 · p. 8

Limitations

  • Evaluation covers question answering on three enterprise-chat benchmarks, leaving benefits for task coordination and long-term planning to be explored.Source: Limitations · p. 9
  • Offline construction is manageable at the evaluated scale, while incremental construction and selective updates are proposed for larger or more frequently updated workspaces.Source: Limitations · p. 9
  • The current formulation focuses on text messages and associated metadata, although images, attachments, and voice messages may contain relevant organizational knowledge.Source: Limitations · p. 9

A.1.1 Answer-Model Robustness

EGCForest remains top-performing across answer-model backbones despite varying baseline rankings, with gains of 13.83–18.52 points over the strongest baseline.

A.1.1 Answer-Model Robustness

  • 13.83–18.52 points over the strongest baseline: EGCForest consistently outperforms memory-system baselines across answer-model backbones.Source: §A.1.1 · p. 11
  • Baseline rankings vary across backbones, but EGCForest remains the top-performing method.Source: §A.1.1 · p. 11
  • Table 4 presents an agent-backbone ablation on GroupMemBench.Source: Table 4 · p. 11

A.2 Dataset Details

The evaluation uses two public enterprise group-chat memory benchmarks and one internal enterprise-chat benchmark; dataset tables summarize benchmark scales and evaluation distributions.

A.2 Dataset Details

  • Evaluation covers two public enterprise group-chat memory benchmarks and one internal enterprise-chat benchmark.Source: §A.2 · p. 11
  • Tables 5, 6, and 7 summarize benchmark scales.Source: §A.2 · p. 11
  • Tables 5, 6, and 7 summarize evaluation distributions.Source: §A.2 · p. 11

A.2.1 EverMemBench

EVERMEMBENCH evaluates long-horizon collaborative memory over enterprise group chats spanning multiple channels, using its official protocol and all released question-answer pairs.

A.2.1 EverMemBench

  • EVERMEMBENCH evaluates long-horizon collaborative memory over enterprise group chats spanning multiple channels.Source: §A.2.1 · p. 11
  • The evaluation uses EVERMEMBENCH's official protocol.Source: §A.2.1 · p. 11
  • The study uses all released EVERMEMBENCH question-answer pairs.Source: §A.2.1 · p. 11

A.2.2 GroupMemBench

The appendix specifies the benchmark sources and question coverage, implementation choices, prompt conventions, and authorship and AI-use disclosure.

A.2.2 GroupMemBench

  • GroupMemBench uses its official corpus and evaluation protocol, with benchmark statistics and question distribution reported in Table 6.Source: §A.2.2 · p. 11, Table 6 · p. 11

A.2.3 Internal Enterprise-Chat Benchmark

  • The internal benchmark uses anonymized enterprise chats to test five retrieval categories, with each question paired to an answer and supporting message evidence.Source: §A.2.3 · p. 11, Table 7 · p. 11

A.3 Implementation Details

  • Evaluation uses Qwen3-Embedding-0.6B for EGCForest and embedding-based reproduced baselines; unavailable prior filtered subsets led reproduced runs to use full public task data.Source: §A.3 · pp. 11–12
  • Forest construction uses chronological k-way grouping with k = 16, while default answering caps final context at 20 messages and the retrieval-budget sweep tests 20–50.Source: §A.3 · p. 12, Table 8 · p. 12

A.4 LLM Prompt Templates

  • Summaries are generated offline and online retrieval uses no LLM calls; open-ended answers must rely only on retrieved memories and track who said what and when.Source: §A.4 · pp. 12–13
  • Offline summaries preserve chronological evidence, attribution, and lifecycle details across Channel, Actor, and Artifact views, while root cards use view-specific fields and omit unsupported information.Source: §A.4 · pp. 12–13

A.5 Ethics Statement

  • The authors state that the research, analysis, and conclusions are human work, with generative AI used only to improve grammar and readability.Source: §A.5 · p. 13
Loading 2610.19836v1…