Persistant memory to LLMs

Aim is to study and strategize ways to implement persistent memory stores to the LLMs.

Persistant memory with Knowledge graphs:

1. Langchain’s LLMGraphTransformer :
   • graph schema, tool calling and prompt tuning
   • Have multiple memory node for different branches ? – easy to define a graph schema
   • eg: Medicine memory node, public sector memory node, Construction, etc
2. Neo4j Graph builder - Online tool

Persitent memory with Vector embedings:

1. Langchain’s long term Agentic memory with LangGraph:
   • Memory component - InMemoryStore ( not clear how it is persistent memory )
   • Types of memory:
     • Semantic memory – uses to store and search factual data (persistent memory) - manage memory tool, search memory tool
     • Episodic memory - Uses human-in-the-loop to refine LLM's ability to perform well ( Few shot examples )
     • Procedure memory - allows the user to update instructions within an agent

Tasks

• Memory pipeline-First version

  A running version of memory pipeline published on AI-Builder.

  Neo4j rendering in AI-Runner:

  Current working solution:

  1. Flask is removed and neo4j exposed on port 8062 directly
  2. neo4j.conf server.bolt.listen_address=0.0.0.0:7687 server.http.listen_address=0.0.0.0:8062
  3. kubectl expose pod 4. DB connection - neo4j://internal_ip:nodeport
  • Streamline the process

• Improvements on memory component:

  Various improvements made on the memory component:

  1. Knowledge graph on "Countries in Europe" - 99 nodes in total

     

  2. Chat Interface - dialog box with "Process":

     

     

  3. Graph generation

     - Batch processing and creating properties to nodes using LLM prompting techniques

     

  4. Graph retrieval

     - Primary Cypher generation based on schema

     - Fallback Cypher generation with fixed schema

     

  5. Stategraph workflow enhancement

     

Innovation LAB 2025

• The memory agent prompt tuning:

  • Detects contradictions against LLM knowledge--> if its a question type , factually wrong information

  • Allows storing factually correct, subjective information

  • Detects contradiction against KG knowledge

  • Video recording

    References

    1. https://medium.com/data-science/building-knowledge-graphs-with-llm-graph-transformer-a91045c49b59
    2. https://learn.deeplearning.ai/courses/long-term-agentic-memory-with-langgraph/lesson/a0k5a/baseline-email-assistant
    3. https://neo4j.com/labs/genai-ecosystem/llm-graph-builder/
    4. https://github.com/neo4j-labs/llm-graph-builder