Video summary

Connecting the Dots with Context Graphs — Stephen Chin, Neo4j

Main summary

Key takeaways

Technology

Summary

Motivation / Problem Statement (AI + Lack of Context)

  • Engineers feel trapped by the AI “revolution”: AI coding/agent tools don’t only assist—they also review and constrain developer work (e.g., PR checks by agents).
  • For critical business decisions, systems built on disconnected enterprise data (e.g., Slack threads, customer tickets, separate systems) often fail because models lack full context.

“Blue Pill vs Red Pill” Framing (Architecture Direction)

  • Blue pill: Stay in a “matrix” of siloed/disparate knowledge, where agents can’t make well-grounded decisions due to missing cross-system context.
  • Red pill: “Escape” by building a system where different data sources are connected, supported by reasoning over:
    • enterprise data sources
    • previous decision traces
    • tool call histories
  • Goal: provide applications with consolidated, contextual memory for better outputs.

Industry Recognition / Momentum

  • Gartner is cited as including context graphs in the AI hype cycle.
  • Foundation Capital is cited as highlighting context graphs as a major startup/industry shift (described as a “$3 trillion startup opportunity”).

Core Technical Approach: Knowledge Graphs + LLMs (Graph RAG / Graph AI)

What Knowledge Graphs Provide

Knowledge graphs are presented as the foundation for context graphs:

  • Store nodes (people, things, companies, relationships)
  • Store relationships with properties
  • Optionally add embeddings (vector information) to support similarity search

How They’re Combined

The talk emphasizes combining:

  • LLMs (language, reasoning, creativity)
  • knowledge graphs (structured context, relationships, explainability)

Retrieval Improvement Example (Medical Grounding)

  • Baseline LLM: answers generically about emphysema care plans.
  • RAG + vector search: adds some patient context but can still produce generic guidance.
  • Graph-grounded retrieval (“grounded complete information”):
    • retrieves richer background (e.g., patient history such as smoking and prior operations)
    • yields more specific recommendations (e.g., medication management, smoking cessation counseling, pulmonary rehab)

Key point: missing context in similarity search can cause overly generic outputs.

Memory Model for Agentic Systems (Three Memory Types)

In the graph, the system distinguishes:

  • Short-term memory: current pipeline state (agent activities, conversation context), stored/persisted in the graph
  • Long-term memory: organized history across long tasks and multiple interactions; requires strong domain modeling and aggregation
  • Reasoning traces (decision provenance): captures not just the result but the why/how, enabling reuse of prior experience; useful for compliance and debugging

Why Graphs Fit “Context Graphs”

  • Relationship-first structure: traversal naturally represents complex connections without heavy relational joins.
  • Performance claims: graph traversal (multi-hop) is highlighted as a major advantage in graph RAG/graph AI.
  • Integration with vectors:
    • use embeddings + vector lookups to find an entry point into the graph
    • use graph algorithms like Louvain for community grouping
  • Explainable decisions and tool support:
    • LLMs can generate Cypher (Neo4j query language)
    • LLMs can create/structure knowledge graphs from unstructured documents
  • Mentioned tooling: Neo4j agent memory package (open-source GitHub) for assembling short-term, long-term, and reasoning memory into a context-graph structure.

Demos

Demo 1: “Lenny Memory Podcast” (Open-Source Project)

  • Purpose: convert dense podcast content into queryable graph context.
  • Flow:
    • loads podcast episodes
    • uses Neo4j agent memory APIs to provide tools for an LLM/agent
  • Example output:
    • extracts locations in episodes
    • builds a navigable map/graph view of mentioned locations
  • Claim: graph format enables a more holistic, dynamically queryable view than partial similarity matches.

Demo 2: Financial Services Context-Graph Application (Agentic Workflow + Auditing)

Architecture overview:

  • Search using context-graph retrieval tools in an agentic architecture
  • combine:
    • knowledge graphs
    • vector search
    • data science algorithms
  • an agent loop writes back into context memory/graph
  • subsequent queries use stored reasoning traces to inform outputs

Open-source application capabilities:

  • integrates with multiple data sources (support ticket system, CRM, internal business data)
  • uses 10 MCP tools (agent-to-tool integrations)
  • uses cloud agents to create OpenAI embeddings and populate the Neo4j context graph
  • provides a UI via Next.js for prompts/questions

Example user question:

  • Whether Jessica Norris should get approval (loan/credit decision).
  • System actions:
    • queries the graph for her history (including margin trades, etc.)
    • shows Cypher queries and graph traversal used for the decision
  • Outputs:
    • recommendation not to approve
    • risk factors and fraud detection patterns, grounded in:
      • prior rejections
      • policy/risk context
      • stored decision rationale
  • Main value: decisions become explainable and auditable because earlier provenance is queryable.

Learning Resources / Tutorial-Style Guidance

  • Neo4j developer education is promoted:
    • a new “context graph” course on GraphAcademy
    • a quick start via a free hosted Aura graph instance (to test before production)
  • Encourages questions after the talk and at the Neo4j booth.
  • Notes upcoming coverage by the next speakers (Zaid and ABK) on more agentic use cases of context graphs.

Main Speakers / Sources

  • Stephen Chin — runs Developer Relations at Neo4j (main speaker)
  • Gartner — cited as officially including context graphs in the AI hype cycle
  • Foundation Capital — cited for a context-graphs industry/opportunity post

Original video