Skip to main content

​
Overview

Neuradex structures AI agent memory using four data types. Each plays a different role, and together they form a powerful memory system.
TypeRoleCharacteristic
EpisodeRaw record of what happenedAccumulated chronologically
TopicEpisode grouping & summaryOptional
KnowledgeConfirmed facts & entitiesAlways reflects latest state
SourceContentDocument chunksMaintains reference to original
How Neuradex is Different: Chunk Search vs Structured MemoryMost RAG services simply split documents into chunks and perform similarity search. Neuradex structures information into meaningful units. By distinguishing “events,” “summaries,” “confirmed facts,” and “originals,” we return the optimal information for your question’s intent.

​
Complex Problems Solved by Structured Memory

Question: “What’s our authentication method?”Simple chunk search:
  • “Using JWT authentication” (doc from 3 months ago)
  • “Planning to switch to OAuth2.0” (meeting notes from last month)
  • “Evaluating auth options” (memo from 6 months ago)
→ Can’t determine which is currentNeuradex:
  • Knowledge: “Authentication method is OAuth2.0” (confirmed fact, latest)
  • Episode: History of changes and reasons
  • SourceContent: Implementation specs
→ Prioritizes confirmed info, provides context when needed

​
Episode

Raw record of “what happened.” Episodes chronologically record all conversations and operation histories. They are the foundation of Neuradex’s memory.

​
What Gets Recorded

  • Conversations: Question and answer exchanges
  • Work: Code modifications, file operations
  • Events: Error occurrences, decisions
  • State changes: Configuration changes, status updates

​
Sources

Episodes are generated from various channels.
SourceDescription
LibrarianConversations with Librarian
ChatbotConversations with chatbots
APIOperations via SDK/API
MCPAI agent operations via MCP server
SlackSlack messages (when integrated)
Episodes are individually searchable. You can look back at past events like “what we discussed last week” or “the situation when an error occurred.” 
"What did I work on yesterday?"
→ Search yesterday's Episodes with time filter

​
Topic

Episode grouping & summary (optional). When multiple related Episodes exist, they are grouped together and a summary is generated. You can understand the gist without reading all Episodes.

​
When Generated

  • Multi-person discussions in Slack
  • Consecutive conversations during meetings
  • Related message groups

​
When Not Generated

  • Single questions
  • Solo work (MCP, etc.)
  • When grouping doesn’t make sense
Not all Episodes need to belong to a Topic. Topics are an optional structure “generated only when needed.”

​
Search

Topics support semantic search. You can discover related topics even with vague queries like “that discussion” or “project progress talk.”

​
Knowledge

Confirmed facts & entities. Knowledge represents “what is true.” If Episodes are records of “what happened,” Knowledge is the state of “how things are now.”

​
Characteristics

  • Always current: Updated when information changes
  • Confirmed info: Definite facts, not ambiguous information
  • Auto-extracted: Automatically generated from Episodes, Topics, and SourceContent

​
Extraction Patterns

Episode: "We decided to use JWT for authentication."
    ↓ Extraction
Knowledge: "Authentication method is JWT"

SourceContent: API specification section
    ↓ Extraction
Knowledge: "API endpoint is /api/v1/auth"

​
Search Priority

For questions, Knowledge is referenced with highest priority. Confirmed Knowledge takes precedence over past Episodes.

​
Knowledge Linking

Knowledge is automatically linked to other Knowledge — no manual linking required. Links include contradiction detection, supersedes (replacement) tracking, and general relationship mapping. This enables automatic context expansion and change history tracking. See the Knowledge API for details on relation types and management.

​
SourceContent

Document chunks. Documents like PDFs, Markdown, and text files are split into chunks (small units) and stored.

​
Characteristics

  • Chunked storage: Large documents split into searchable sizes
  • Original reference: Maintains reference to source document
  • Vector search: Searchable by semantic similarity

​
Supported Formats

  • PDF
  • Markdown
  • Text files
  • Other document formats

​
Data Flow

The four data types are interrelated and unified in a vector collection.

​
Next Steps

Quickstart

Create your first knowledge

Knowledge API

Manage knowledge with SDK