Dynamic Memory Forest: Precisely Tracking Long‑Range Dialogue Trajectories for Highly Coherent Responses

Problem Statement

Despite the expansion of large‑model context windows to the million‑token scale, open‑domain conversations that span long periods remain fragmented, causing models to lose deep semantic links across turns. Existing approaches—summarization compression or fragmented similarity retrieval—break essential bridge information and fail to connect high‑entropy “bridge nodes” that link multiple topics.

Inspiration from Cognitive Science

Human memory follows a consolidation‑then‑growth process. By clustering dialogue utterances into thematic blocks and establishing cross‑topic associative networks via synaptic plasticity, the authors propose the Dynamic Memory Forest (DMF) to emulate this mechanism.

Dynamic Memory Forest Framework

DMF converts disordered long‑term dialogue into a structured “memory forest” composed of multiple memory trees. The framework consists of four stages:

Stage 1 – Memory Consolidation : After a conversation ends, utterance embeddings are reduced with UMAP and clustered softly using a Gaussian Mixture Model (GMM). Each utterance receives a thematic entropy score; high‑entropy utterances become candidate bridge nodes.

Stage 2 – Memory Growth : Consolidated units are ordered by timestamps to form dialogue chains. An LLM generates a concise summary for the chain’s root node, creating a memory tree. A Group Relative Voting Optimization (GRVO) mechanism decides whether a new tree should remain independent or graft onto an existing tree based on cosine‑similarity votes exceeding a threshold.

Stage 3 – Entropy‑Driven Memory Walk (EDMW) : When a user query arrives, EDMW searches the forest for an optimal memory path, balancing global relevance and local coherence while amplifying high‑entropy bridge nodes to enable cross‑tree jumps.

Stage 4 – Response Generation : Visited nodes are sorted by global timestamps; root‑node summaries are injected as hierarchical memory cues before feeding the assembled context to the LLM, which produces a coherent reply.

Experimental Evaluation

Experiments were conducted on three real‑world open‑domain long‑dialogue datasets—Conversation Chronicles (CC), Multi‑Session Chat (MSC), and GapChat (GC)—using GPT‑4o (128K) and Gemini 2.5 (1M) as base models. Baselines covered four dominant paradigms: Full History, Structured RAG (e.g., MemTree), Agentic Memory (e.g., MemoryOS), and Memory‑Enhanced Generation (e.g., THEANINE).

Automatic metrics (including BLEU, ROUGE, and Mauve) show DMF achieving the best scores across all datasets. Notably, the Mauve score, which measures similarity to human style, improves dramatically, demonstrating that structured dialogue trajectories are more effective than merely feeding ultra‑long contexts.

Case Study

A user asks for advice on a “new oven.” DMF first locates the “baking” tree node (A1), then discovers a high‑entropy bridge node linking to “healthy food” (C2). The walk jumps to the “fitness” tree (B2) and finally to the core theme of weight loss (B1), enabling the model to suggest “sugar‑free cookies” that respect both the new oven and the user’s diet goal.

Efficiency Analysis

Compared with methods that rebuild trees frequently (e.g., MemTree), DMF processes conversations in batch after each session, resulting in the lowest total token consumption among dynamic memory approaches. Construction time and retrieval latency are also significantly reduced, supporting real‑time, long‑range interactions.

Conclusion

The Dynamic Memory Forest successfully applies cognitive‑science principles to LLM memory architecture, addressing fragmentation in long‑term dialogue through clustering‑consolidation, grafting‑growth, and associative walking. The results confirm that organized memory structures and cross‑topic trajectories are far more critical than merely expanding context windows for achieving coherent, human‑like long‑range conversations.