Data Model

MemNexus stores all data in a Neo4j graph database. Understanding the graph schema helps you write better queries and build more effective integrations.

Node types

Memory

The fundamental unit of storage.

Property	Type	Description
id	UUID	Unique identifier
content	string	The memory content
memoryType	string	Type (default: episodic)
importance	float	0-1 significance score
embedding	float[]	1536-dimension OpenAI vector
createdAt	datetime	When the memory was created
updatedAt	datetime	When last modified

Conversation

Groups related memories into logical sessions.

Property	Type	Description
id	UUID	Unique identifier
title	string	Conversation title
createdAt	datetime	When started
updatedAt	datetime	When last memory was added

Topic

Labels that emerge from memories for categorization.

Property	Type	Description
id	UUID	Unique identifier
name	string	Topic name
embedding	float[]	Topic vector for similarity

Fact

Structured knowledge as subject-predicate-object triples.

Property	Type	Description
id	UUID	Unique identifier
subject	string	Fact subject
predicate	string	Relationship type
object	string	Fact object
confidence	float	0-1 confidence score

Entity

People, tools, projects, and concepts that appear across memories.

Property	Type	Description
id	UUID	Unique identifier
name	string	Entity name
type	string	Entity type (person, tool, project, concept)

Community

Groups of related topics detected via graph algorithms.

Property	Type	Description
id	UUID	Unique identifier
name	string	Community name
algorithm	string	Detection algorithm used

Relationships

Memory  ──[HAS_TOPIC]──────→ Topic
Memory  ──[MENTIONS]────────→ Entity
Memory  ──[BELONGS_TO]──────→ Conversation
Memory  ──[RELATED_TO]──────→ Memory
Fact    ──[ABOUT]───────────→ Entity
Fact    ──[HAS_SUBJECT]─────→ Entity
Fact    ──[HAS_OBJECT]──────→ Entity
Topic   ──[RELATED_TO]──────→ Topic
Topic   ──[BELONGS_TO]──────→ Community
Entity  ──[RELATED_TO]──────→ Entity

Indexes

Vector indexes

MemNexus creates vector indexes on Memory.embedding for semantic search using cosine similarity. Embeddings are generated using OpenAI's text-embedding-3-small model (1536 dimensions).

Full-text indexes

Full-text indexes are maintained on Memory.content for keyword search. These support the keyword component of hybrid search.

Standard indexes

Standard B-tree indexes on:

• Memory.id, Memory.createdAt
• Conversation.id
• Topic.id, Topic.name
• Fact.id
• Entity.id, Entity.name

Search pipeline

When you search, MemNexus runs:

1. Semantic search — Embeds your query, finds nearest neighbors in vector space
2. Keyword search — Full-text search across memory content
3. Reciprocal Rank Fusion — Merges and re-ranks results from both searches

This hybrid approach means both "that deployment issue" (semantic) and "ECONNREFUSED" (keyword) return relevant results.