Build Daily
Tinley Park · August 23, 2026What’s Next for AI Agents? Memory, Context, and Graphs
A speaker at the Neo4j Graph Gathering discusses the future of AI agents, focusing on agentic memory, context graphs, and the evolution from RAG to production-ready systems that can take action and self-improve.
The outline
Tap a timestamp to jump there in the video.
Introduction to Neo4j Graph Gathering
The speaker expresses enthusiasm for attending the Neo4j Graph Gathering and learning from others about different technologies.
Agentic Memory as a Key Database Topic
The speaker identifies agentic memory as the next hot topic for databases, noting that current agent memory is often lost when sessions end.
Context Graphs for Agent Improvement
The speaker highlights context graphs as a tool for tracking agentic decisions, enabling self-learning, and improving personalization of recommendations.
MCP Toolbox for Databases Overview
As technical lead for the MCP Toolbox for databases, the speaker describes its current focus on build-time tools for connecting to over 40 different databases.
Expanding into Runtime Memory Tools
The team is interested in developing memory tools that augment both developer assistance and runtime end-user applications for personalized results.
Learning from Graph Database Experts
With expertise in traditional databases, the speaker values learning how graph databases can deeply embed relationships between data sources to augment AI agents.
Evolution Beyond RAG for Production Agents
The speaker is focused on best practices for connecting data with AI agents and understanding how the field has evolved from RAG to production systems that can take action and self-improve.
Connecting Technical Expertise to Business Needs
The speaker emphasizes the importance of connecting with people in different roles to understand business use cases and relate deep technical expertise to customer problems.
Closing a laptop often clears an agent's working context, so durable memory becomes the difference between demos and production systems. For solo builders shipping AI products, treating memory as infrastructure—not just prompt state—lets agents act on accumulated history instead of starting over each run.