Context graphs connect enterprise data, previous decisions, tool calls, and reasoning so agents can use the relationships behind information.
2
Graph-based retrieval can give healthcare recommendations grounded in a patient's history instead of generic advice from a vector search.
3
Reasoning traces let people inspect why an agent made a decision, which supports later decisions, debugging, and compliance.
Summary
Stephen Chin argues that retrieval systems lose important context when enterprise information is split across Slack, customer threads, and business systems. A vector RAG system may find relevant text, but it can miss the relationships between a patient's history, earlier decisions, policies, and outcomes. Knowledge graphs connect those entities and can combine graph traversal with embeddings and language models. Chin describes three memory types: short-term memory for the current agent workflow, long-term memory for domain history, and reasoning memory for the decisions and tool calls that produced an outcome. His financial services demo shows an agent assessing a loan while exposing a customer's related trades, a previous rejection, risk factors, and fraud patterns. The human still makes the final decision, but can see the information and reasoning behind the recommendation.
Disconnected enterprise data leaves agents without the context needed for business decisions
Chin describes engineers as stuck with knowledge spread across Slack discussions, customer threads, and separate enterprise systems. When an agent has to make a critical decision from that scattered information, it cannot form a complete view. His proposed alternative connects enterprise data sources, previous decision traces, and tool calls into a system of reasoning. He frames this as moving from isolated pieces of knowledge toward a structure where applications can retrieve the relationships between facts and decisions.
Knowledge graphs connect entities and can combine structured relationships with embeddings
A knowledge graph stores nodes such as people, things, and companies, along with the relationships between them and their properties. Chin uses an example where people, a car, and their connections are represented together, with embeddings added to support similarity searches. He says language models contribute language, reasoning, and creativity, while knowledge graphs contribute knowledge, context, and enrichment. Together, they can store relationships, visualize relevant data, find patterns, and support further analysis.
Graph retrieval produces patient-specific care recommendations that vector search can miss
In Chin's healthcare example, a baseline language model knows what emphysema is and gives broad advice. A vector RAG system adds some patient information, but still recommends generic activities such as respiratory therapy, deep breathing, and coughing exercises. A graph-grounded system brings in the patient's diagnosis, operations, and other history. It recommends medication management, smoking cessation counseling, and pulmonary rehabilitation exercise because it knows the patient smokes and has had an operation. The difference is the connected background information available to retrieval.
Agents need short-term, long-term, and reasoning memory
Chin separates agent memory into three forms. Short-term memory contains the current conversation and the state of the active workflow. Long-term memory stores information from longer tasks, repeated procedures, and multiple customer interactions, supported by a domain model for entities and business processes. Reasoning traces record how the agent reached a result, including earlier decisions and the experience behind them. Those traces can make decisions repeatable and provide material for compliance and debugging.
Graph relationships make agent memory queryable and explainable
Chin says relationships are first-class in knowledge graphs, so applications can traverse connected information without assembling a series of table joins. Graph embeddings can provide vector lookups as an entry point, while algorithms such as Louvain can group related communities. The resulting structure lets applications inspect how information connects and produce decisions with an explanation. He points to the open-source Neo4j agent memory package, which brings short-term memory, long-term memory, and reasoning memory into a context graph.
Context graphs preserve why a decision was made, rather than only recording the result
A traditional audit log may record an outcome, while a context graph records the decision traces created while models were evaluated. It organizes those traces around entities and relationships and brings together information that might otherwise remain in Slack, email, or informal discussions. An agent can retrieve earlier decisions and recommendations, then add its own recommendation to the reasoning history for later queries. Chin describes an architecture that retrieves from graphs, vectors, and data science algorithms, then writes the agent's new context back into the graph.
The financial services demo exposes the evidence behind a loan recommendation
The demo models people and organizations, along with decisions, transactions, approvals, policies, risk factors, and employee reasoning. It connects a support ticket system, a CRM, and an internal business data system through MCP tools, then loads the resulting information into a Neo4j context graph. When assessing Jessica Norris, the application retrieves her history and related margin trades, shows the Cypher queries, and displays the graph traversal. It also finds a previous rejection and fraud detection patterns. The model recommends denying the loan, while the visible evidence gives the human decision-maker reasons they can review and justify.
"And as a human who's using the system to make decisions, this is the sort of information you need to actually make a decision you can stand behind and you can justify to your organization."15:21
Who should watch
You are building an agent that must use information spread across business systems and need a way to connect that information to the entities and decisions it concerns.
Your RAG application returns relevant text but loses patient, customer, transaction, or policy history that changes the answer.
You need people to inspect an agent's evidence and reasoning before accepting a recommendation, especially in regulated or high-risk workflows.
