feat(memory): introduce memory module

[Appointat]

Background

Chat2Graph is dedicated to building a powerful Multi-Agent System (MAS) that aims to enhance the intelligence of conversational interactions through the deep integration of graph technology and artificial intelligence. In advanced agent systems, the Memory module is an indispensable core component, providing agents with the foundation to learn, adapt, and offer more context-aware and personalized services.

Current Status

Currently, Chat2Graph achieves a degree of short-term information retention through mechanisms like message history, workflows (including evaluator), and an automated planner. However, to realize more advanced intelligent behaviors such as cross-session information retention, complex environment understanding, insight generation, and an experience pool, we urgently need to develop or introduce a dedicated, well-designed memory module.

Desired Features

We expect to introduce an independent memory module responsible for:

1. Multi-level Memory Management: Support for short-term memory (e.g., current conversation context, temporary calculation results) and long-term memory (e.g., user preferences, historical interaction summaries, learned knowledge).
2. Automated Memory Operations: Provide flexible memory storage, fast and accurate retrieval, timely updates, and a reasonable forgetting mechanism.
3. Seamless Integration: Ability to tightly integrate with existing core modules like Leader&Expert, Operator, Reasoner, providing memory support for their decision-making and execution.
4. Scalability and Configurability: Allow for future expansion of memory types and storage structures according to needs, and support system-level memory module configuration.

Value

Introducing a memory module will bring the following values to Chat2Graph:

• Enhance Agent's Long-term Learning and Adaptation Capabilities
• Improve Contextual Understanding and Personalized Interaction
• Support More Complex Reasoning and Planning
• Increase Multi-Agent Collaboration Efficiency

Preliminary Ideas

We preliminarily envision the memory module as a DIKW memory model.

To support the DIKW layered model, this module may need to develop the following sub-modules/components:

• Memory Storage: Consider using a combination of vector databases, graph databases, and traditional databases to meet the storage needs of different types of memory.
• Memory Processing: Including algorithms for memory encoding, compression, summarization, association, and forgetting.
• Memory Interface: Provide standardized memory access APIs for other modules.