Memory Strategies based on design patterns with connector Types

[marcellodesales]

What you'd like to see

• The current memory model uses sqlite to implement the memory capability of agents
• What if I want to implement the concept of a Knowledge Graph?
• Maybe implement a specification such as mem0.ai's approach

Tip

The "3x improvement" associated with Neo4j's GraphRAG (Graph Retrieval-Augmented Generation) refers to studies showing that this approach, on average, improves the accuracy of LLM responses by three times compared to traditional, vector-only RAG methods. It is not a specific "agent" but rather an architectural approach and set of tools that leverage knowledge graphs to provide superior context and reasoning capabilities to AI agents.

• 3x improvent on Agents Brain capability when using a GraphRag (https://neo4j.com/blog/genai/graphrag-manifesto/#:~:text=%231:%20Higher%20Accuracy%20&%20More,backed%20by%20a%20knowledge%20graph.)

Why you'd like to see it

• Memory should be a plugable data type that instructs the Cagent to choose the approach to "Rag".
• The developer, knowing different strategies and design patterns around Rag implementation, should be able to choose the memory strategy at minimum
• Long-term memory: I think this is currently used as a Rag in the current implementation
• Short-term memory: Using Whiteboard Memory for Short-Term Context, where a multi-agent system exchanges information through a whiteboard.

This is discussed in detail by different vendors. I'd like to call out two:

• https://learn.microsoft.com/en-us/semantic-kernel/frameworks/agent/agent-memory
• https://neo4j.com/blog/developer/modeling-agent-memory/

What I would like to see is a high-level construct that defines the shared short-term memory model for agents, where types are implementations following the concept of "docker plugins"... If cagent plugins could be one of the architectural ways to deliver on these features, I think developers would benefit from the community-driven solutions.

# First class, top-level construct for a short-term memory
memory:
  name: whiteboard
  kind: neo4j-graph
  metadata:
    url: localhost:4499
    auth: secret-1234

Workarounds?

• Messaging passing among the agents throught the design of the memory construct.
• Major problem is in agent versioning, memory structure versioning, etc.

Additional context