Short-Term Memory

In the previous lesson, you learned why short-term memory is necessary — without it, every prompt reaches the LLM cold, with no knowledge of what came before.

In this lesson, you will learn how short-term memory is modeled as a graph, what properties each message node stores, and what the library does automatically when you add a message.

Understanding the characteristics of short-term memory

Short-term memory has three defining characteristics that distinguish it from the other memory layers.

Characteristic	Short-Term Memory
Lifetime	Ephemeral — exists for the duration of a session
Access pattern	Sequential — recent messages first
Primary question	"What did you discuss?"

Characteristic

Short-Term Memory

Lifetime

Ephemeral — exists for the duration of a session

Access pattern

Sequential — recent messages first

Primary question

"What did you discuss?"

Examining the graph schema

A Conversation node acts as the entry point and links to its messages through a chain of FIRST_MESSAGE and NEXT_MESSAGE relationships.

mermaid

Short-term memory schema

graph LR
    C([Conversation \n id, user_id, title]) -->|FIRST_MESSAGE| M1([Message \n role: user \n embedding])
    M1 -->|NEXT_MESSAGE| M2([Message \n role: assistant])
    M2 -->|NEXT_MESSAGE| M3([Message \n role: user])

Each Message node stores:

role — user, assistant, or system
content — the message text
timestamp — when the message was created
session_id — which session the message belongs to
embedding — a vector used for semantic search across message history

What happens automatically when a message is added

When you add a message to short-term memory, the library performs these steps automatically:

Creates the Message node and links it into the Conversation chain
Runs the entity extraction pipeline on the message content
Promotes discovered entities into the long-term memory layer
Links the message to any extracted entities using MENTIONS relationships

This automatic extraction connects short-term and long-term memory within the same graph.

Summary

In this lesson, you learned how short-term memory works as a graph:

Conversation → Message chain — short-term memory stores the session’s working context as a linked list of nodes
Entity extraction — each message is automatically processed and discovered entities are promoted to long-term memory
Vector embeddings — each Message node stores an embedding to support semantic search across message history

In the next lesson, you will learn the short-term memory API methods and the Cypher each one generates.

Context Graphs: Agent Memory with Neo4j

Introduction to Agent Memory