Abstract:The application of large language models to facilitate automated software operations and tool generation (SOTG), thus augmenting software productivity, mirrors the early stages of human evolution when the ability to create and use tools accelerated the progress of civilization. These complex tasks require AI to continuously summarize and improve. Current research often overlooks the importance of converting real-time task experiences into system memory and differentiating the value of existing knowledge for future reference. This paper addresses these issues by evolving external memory models into Memory-Loop Networks for timely memorization and experience referencing. We also enhance a RAG mechanism with knowledge precision segmentation to utilize memory based on value differentiation, and design the MaxMind model for SOTG <a class="link-external link-http" href="http://accordingly.To" rel="external noopener nofollow">this http URL</a> demonstrate our approach, we developed MaxMind4Sheet, an electronic spreadsheet processing system aligned with the MaxMind philosophy. Comparative experiments with SheetCopilot have demonstrated that the accumulation and recycling of task memories lead to a steady enhancement in task success rate, with an improvement rate of approximately 3%-6% per round in this implementation example. Note that as the memories continue to grow, this cumulative improvement may be substantial. The inclusion of memory recycling can also boost the system's task execution efficiency by up to 25%, and it can address the retraining issue faced by LLMs when handling specialized tasks through memories transfer.These suggest that MaxMind has significant potential to enhance the capabilities and productivity of LLM systems in SOTG.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are: how to enhance the performance of large - language models (LLMs) in software operation and tool generation (SOTG) tasks by introducing an external memory mechanism, thereby improving software productivity. Specifically, the paper aims to solve the following problems: 1. **Accumulation of real - time task experience and update of system memory**: - Current research often ignores the importance of transforming real - time task experience into system memory, resulting in the system being unable to effectively use past experiences for improvement. 2. **Differentiated treatment of knowledge value**: - Existing methods fail to distinguish the value of existing knowledge when dealing with it, affecting the decision - making efficiency and accuracy of the system. 3. **Dynamic adaptation and continuous improvement**: - The performance of existing frameworks declines when facing complex or ambiguous tasks, and they lack the ability to be retrained, which limits their application effects in specific tasks. To solve these problems, the paper proposes a new concept - **Memory - Loop Network (MLN)**, and realizes this concept by designing the MaxMind model. MaxMind can dynamically update the external knowledge base by introducing a memory - loop mechanism, and adjust the precision and relevance of memory according to task requirements, thereby improving the reasoning ability and execution efficiency of the system. ### Main contributions 1. **New Memory - Loop Network (MLN)**: - The output of MLN serves as a reference basis for subsequent request inputs, forming a feedback loop, so that the output can be re - used as input to dynamically update the memory slot. This upward cycle enhances the system's ability to learn from operational experiences and continuously improve. 2. **Adaptive RAG method based on knowledge value assessment**: - A new RAG method is proposed, which assigns different precision levels according to the value assessment of knowledge, optimizing the efficiency and precision of RAG. 3. **Prototype implementation and experimental verification**: - Based on the llama3.1 - 70B model and vector database, the MaxMind4Sheet prototype system is implemented, and its performance and scalability are verified through a series of carefully designed experiments. Through these innovations, MaxMind significantly enhances the capabilities of LLM systems in SOTG tasks, especially in terms of task - memory accumulation and reuse, showing great potential.

MaxMind: A Memory Loop Network to Enhance Software Productivity based on Large Language Models

MemoryBank: Enhancing Large Language Models with Long-Term Memory

Memory Sharing for Large Language Model based Agents

Empowering Working Memory for Large Language Model Agents

LDM$^2$: A Large Decision Model Imitating Human Cognition with Dynamic Memory Enhancement

Dual Memory Model for Experience-Once Task-Incremental Lifelong Learning.

An Artificial Neuron for Enhanced Problem Solving in Large Language Models

Think-in-Memory: Recalling and Post-thinking Enable LLMs with Long-Term Memory

MEMORYLLM: Towards Self-Updatable Large Language Models

HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model

FastMem: Fast Memorization of Prompt Improves Context Awareness of Large Language Models

Enhancing Large Language Model with Self-Controlled Memory Framework

Augmenting Language Models with Long-Term Memory

RecallM: An Adaptable Memory Mechanism with Temporal Understanding for Large Language Models

UniMem: Towards a Unified View of Long-Context Large Language Models

From Isolated Conversations to Hierarchical Schemas: Dynamic Tree Memory Representation for LLMs

MemLong: Memory-Augmented Retrieval for Long Text Modeling

MemSim: A Bayesian Simulator for Evaluating Memory of LLM-based Personal Assistants

Unlocking Memorization in Large Language Models with Dynamic Soft Prompting

RAM: Towards an Ever-Improving Memory System by Learning from Communications