What is Agent Memory? - William Lyon

How can we build AI agents that actually remember and learn? Explore a new graph-based approach that dramatically cuts token usage and improves reasoning.

#1about 1 minute

What is Neo4j and its graph intelligence platform

Neo4j has evolved from an embedded graph database into a scalable graph intelligence platform for real-time use cases like fraud detection.

#2about 3 minutes

Using real-time graphs to improve AI personalization

Real-time graph traversal provides more relevant, personalized recommendations by avoiding the stale data issues common in batch processing systems.

#3about 3 minutes

Defining AI agents as a loop with tools

An AI agent is fundamentally a reasoning loop that uses a set of tools to interact with its environment and achieve a specific goal.

#4about 4 minutes

The importance of human-in-the-loop for agents

Fully autonomous agents are often impractical; a human-in-the-loop approach is crucial for augmenting knowledge work and ensuring compliance in regulated industries.

#5about 4 minutes

The challenge of agentic memory and user preferences

Agents need memory to learn user preferences and avoid repetitive interactions, but implementing a scalable memory layer presents significant systems challenges.

#6about 4 minutes

Neo4j's three types of graph-based agent memory

The memory model is structured into short-term (conversations), long-term (a domain-driven knowledge graph), and reasoning memory (tool call traces).

#7about 5 minutes

How reasoning memory enables efficient tool selection

By recording and evaluating the outcomes of past reasoning paths, agents can learn which tool call sequences are most effective for similar future tasks.

#8about 4 minutes

Getting started with the Neo4j Agent Memory package

Developers can integrate graph-based memory into Python agent frameworks like LangChain and CrewAI using a simple package and the Neo4j Aura free tier.

#9about 3 minutes

The three-stage pipeline for entity extraction

A composable pipeline extracts entities efficiently by using statistical NLP and small local models first, reserving expensive LLMs as a final fallback.

#10about 3 minutes

Measuring success with improved accuracy and efficiency

The primary success metric for adopting graph-based memory is a dramatic increase in accuracy, which then paves the way for optimizing efficiency.

#11about 3 minutes

The future of agent memory is shared conventions

The goal is to establish a language-agnostic shared memory substrate that allows agents built in different frameworks to collaborate using the same data model.

#12about 2 minutes