D-Mem: A Dual-Process Memory System for LLM Agents
arXiv:2603.18631v1 Announce Type: new Abstract: Driven by the development of persistent, self-adapting autonomous agents, equipping these systems with high-fidelity memory access for long-horizon reasoning has emerged as a critical requirement. However, prevalent retrieval-based memory frameworks often follow an incremental processing paradigm that continuously extracts and updates conversational memories into vector databases, relying on semantic retrieval when queried. While this approach is fast, it inherently relies on lossy abstraction, frequently missing contextually critical information and struggling to resolve queries that rely on fine-grained contextual understanding. To address this, we introduce D-Mem, a dual-process memory system. It retains lightweight vector retrieval for routine queries while establishing an exhaustive Full Deliberation module as a high-fidelity fallback. To achieve cognitive economy without sacrificing accuracy, D-Mem employs a Multi-dimensional Quali
arXiv:2603.18631v1 Announce Type: new Abstract: Driven by the development of persistent, self-adapting autonomous agents, equipping these systems with high-fidelity memory access for long-horizon reasoning has emerged as a critical requirement. However, prevalent retrieval-based memory frameworks often follow an incremental processing paradigm that continuously extracts and updates conversational memories into vector databases, relying on semantic retrieval when queried. While this approach is fast, it inherently relies on lossy abstraction, frequently missing contextually critical information and struggling to resolve queries that rely on fine-grained contextual understanding. To address this, we introduce D-Mem, a dual-process memory system. It retains lightweight vector retrieval for routine queries while establishing an exhaustive Full Deliberation module as a high-fidelity fallback. To achieve cognitive economy without sacrificing accuracy, D-Mem employs a Multi-dimensional Quality Gating policy to dynamically bridge these two processes. Experiments on the LoCoMo and RealTalk benchmarks using GPT-4o-mini and Qwen3-235B-Instruct demonstrate the efficacy of our approach. Notably, our Multi-dimensional Quality Gating policy achieves an F1 score of 53.5 on LoCoMo with GPT-4o-mini. This outperforms our static retrieval baseline, Mem0$^\ast$ (51.2), and recovers 96.7\% of the Full Deliberation's performance (55.3), while incurring significantly lower computational costs.
Executive Summary
This article introduces D-Mem, a dual-process memory system designed to address the limitations of prevalent retrieval-based memory frameworks. By combining lightweight vector retrieval with an exhaustive Full Deliberation module, D-Mem achieves cognitive economy without sacrificing accuracy. The Multi-dimensional Quality Gating policy dynamically bridges these two processes, making it an effective solution for high-fidelity memory access in long-horizon reasoning. The experimental results demonstrate the efficacy of D-Mem, outperforming a static retrieval baseline and recovering 96.7% of the Full Deliberation's performance while incurring significantly lower computational costs.
Key Points
- ▸ D-Mem is a dual-process memory system that combines lightweight vector retrieval with an exhaustive Full Deliberation module.
- ▸ The Multi-dimensional Quality Gating policy dynamically bridges these two processes to achieve cognitive economy without sacrificing accuracy.
- ▸ D-Mem outperforms a static retrieval baseline and recovers 96.7% of the Full Deliberation's performance while reducing computational costs.
Merits
Strength in Addressing Lossy Abstraction
D-Mem's ability to retain high-fidelity memory access addresses the limitations of prevalent retrieval-based memory frameworks that rely on lossy abstraction.
Cognitive Economy without Sacrificing Accuracy
The Multi-dimensional Quality Gating policy enables D-Mem to dynamically switch between vector retrieval and Full Deliberation, achieving cognitive economy without compromising accuracy.
Demerits
Computational Costs
The computational costs of D-Mem may be higher than those of static retrieval-based memory frameworks, although the article suggests that D-Mem incurs significantly lower costs than Full Deliberation.
Complexity
D-Mem's dual-process architecture may add complexity to the design and implementation of memory systems, which could be a limitation for some applications.
Expert Commentary
The introduction of D-Mem marks a significant advancement in the development of memory systems for long-horizon reasoning. By addressing the limitations of prevalent retrieval-based memory frameworks, D-Mem offers a more effective solution for high-fidelity memory access. The experimental results demonstrate the efficacy of D-Mem, and its potential applications in various fields make it an exciting area of research. However, the computational costs and complexity of D-Mem may need to be carefully evaluated in different contexts. Overall, D-Mem is a valuable contribution to the field of artificial intelligence and memory systems.
Recommendations
- ✓ Future research should focus on further optimizing the computational costs of D-Mem and exploring its applications in various fields.
- ✓ Developers should consider incorporating D-Mem into their memory systems to improve the performance and accuracy of autonomous agents.
