EGCForest: Building a Structured Agent Memory Framework for Enterprise Group Chats

Title

EGCForest: Building a Structured Agent Memory Framework for Enterprise Group Chats

Meta Description

EGCForest proposes a three-forest architecture that organizes immutable message logs from enterprise group chats into three dimensions—channel, actor, and artifact—and performs joint retrieval through cross-view links, improving answer accuracy by up to 20.85 points over MemOS on three enterprise group chat benchmarks.

Article Content

Enterprise group chats contain a wealth of organizational information—project progress, key decisions, task ownership, and more. For building enterprise-grade AI agents, this data should be a valuable memory asset. But the reality is that reliably retrieving information from group chats is extraordinarily difficult: evidence is scattered across interleaved discussion threads, different channels, and long time spans, while the meaning and authority of information depend heavily on who said it. Traditional approaches—whether RAG's independent chunk retrieval or memory systems' summary-based consolidation—struggle to reliably recall such distributed evidence.

Why Existing Approaches Fall Short for Enterprise Group Chat Retrieval

RAG works by splitting documents into independent chunks and recalling them by similarity. This logic performs well within a single document, but enterprise group chat conversations are inherently multi-party, interleaved, cross-channel, and cross-temporal. The owner of a task might appear in the product channel on Monday, while the acceptance criteria are buried in an engineering thread on Wednesday. Independent retrieval cannot establish these cross-fragment associations.

Memory systems typically generate conversation summaries and then consolidate them. Summaries lose details—especially the semantic differences introduced by speaker identity. The same phrase, "let's rethink this proposal," means something entirely different when said by a technical lead versus an intern. Summary-based memory flattens out this identity-conditional information, leaving the agent unable to judge which information truly carries decision-making weight.

RAG (Retrieval-Augmented Generation) is the current mainstream retrieval approach: it splits documents into independent chunks, uses vector similarity to recall the fragments most relevant to a query, and then hands them to a large model to generate an answer. Its core assumption is that knowledge exists as relatively independent document fragments, and that the retrieval unit is the semantic unit. MemOS, on the other hand, is a memory operating system for agents that compresses conversation history into summaries and consolidates them hierarchically, managing long-term context in a way closer to human memory. The shared limitation of both is this: RAG lacks structural associations across chunks and cannot reconstruct the distributed chain of "who said what, when, and corresponding to which deliverable"; MemOS's summary compression loses fine-grained conditional information such as speaker identity. EGCForest is designed precisely to address these two deficiencies, replacing implicit compression with explicit structure.

EGCForest's Three-Forest Architecture

The core idea of EGCForest is to reorganize immutable message logs into three complementary forest structures, each corresponding to one of three critical contextual dimensions in enterprise group chats.

Channel forest organizes workflows along the timeline, reconstructing the progression of tasks within each channel. Actor forest builds context centered on speakers, preserving the critical condition of "who said what." Artifact forest tracks the lifecycle of deliverables—the complete chain from proposal through revision to final confirmation.

The three forests do not exist in isolation. EGCForest establishes cross-view connections through a Placement Map and sparse structural bridges. This means that when retrieving a task, the system can simultaneously gather evidence along the timeline, speaker threads, and deliverable paths—rather than gambling on a single dimension.

Retrieval and Assembly Mechanism

On top of the memory skeleton, EGCForest's retrieval strategy jointly leverages three types of signals: message-level relevance, hierarchical context from forest summaries, and cross-view links.

This joint retrieval design targets precisely the "distributed evidence" problem. Traditional RAG only looks at message-level similarity, whereas EGCForest, after recalling a single message, can also traverse the forest structure to pull in its parent hierarchical summaries and cross-view associated nodes, reassembling scattered evidence into complete context.

Sparse structural bridges refer to establishing only a small number of explicit connections between key nodes, rather than full cross-linking. This design keeps the cost of cross-view jumps manageable and avoids structural explosion. Message-level relevance is typically computed via vector or keyword similarity and is responsible for quickly locating candidate messages; hierarchical context comes from forest summaries, providing background such as the channel and task stage a message belongs to; cross-view links connect fragments of the same task across the three dimensions of timeline, speaker, and deliverable. After joint retrieval of all three, the system performs "reassembly": using recalled messages as anchors, it pulls in upstream and downstream evidence along structural bridges, assembles them into self-consistent context units, and then hands them to the model for answering—thereby mitigating the evidence fragmentation caused by local recall.

Experimental Results: Up to 20.85 Points Improvement Over MemOS

The paper conducted experiments on three enterprise group chat benchmarks. The results show that EGCForest consistently outperforms RAG and memory system baselines, achieving up to a 20.85 percentage point improvement in answer accuracy over MemOS, while also supporting efficient online retrieval.

This margin of improvement is noteworthy. A gap of 20.85 points indicates that structured multi-view organization genuinely brings a qualitative change in distributed evidence recall—not merely a gain at the level of engineering optimization.

Answer accuracy is a core metric on enterprise group chat retrieval benchmarks, typically calculated as the match between model answers and ground truth or via human/model evaluation, measuring whether an agent can reach correct conclusions based on recalled evidence. A 20.85 percentage point improvement means that under identical model and prompt conditions, EGCForest recalls significantly higher-quality evidence—not by relying on a stronger generative model, but through structured evidence organization. Online retrieval efficiency refers to the system's ability to complete multi-view joint recall under real-world query latency constraints, demonstrating that the indexing overhead of structural bridges and Placement Map is manageable and that the approach is viable for production deployment. Results that simultaneously deliver on both quality and latency are a key signal for judging whether a memory layer design can enter production environments.

Implications for Enterprise Agent Deployment

EGCForest solves a long-underestimated problem: enterprise group chats are not ordinary documents—they are organizational memory with roles, weights, and workflow structure. Treating this kind of data as flat text inevitably leads to information loss.

The three-forest architecture offers a viable structured path, enabling agents to distinguish the temporal脉络, speaker conditions, and deliverable states of information during retrieval. For teams building enterprise-grade agents, this framework is worth considering as a reference direction for memory layer design.

One-line Summary

EGCForest reorganizes enterprise group chats into a three-forest structure to enable distributed evidence retrieval, improving accuracy by up to 20.85 points over MemOS.

Paragraph Summary

Enterprise group chats contain organizational information such as project progress, decisions, and task ownership, but because discussions are interleaved, cross-channel, cross-temporal, and semantically dependent on speaker identity, traditional RAG's independent chunking and memory systems' summary consolidation both struggle to reliably recall distributed evidence. EGCForest reorganizes immutable message logs into three complementary forests—Channel, Actor, and Artifact—corresponding to the three contextual dimensions of timeline, speaker, and deliverable, and establishes cross-view connections through a Placement Map and sparse structural bridges. During retrieval, it jointly leverages message-level relevance, hierarchical context from forest summaries, and cross-view links to reassemble scattered evidence into complete context. On three enterprise group chat benchmarks, its answer accuracy improves by up to 20.85 percentage points over MemOS while also supporting efficient online retrieval, providing a structured reference for memory layer design in enterprise-grade agents.