Core Concepts

Understanding the Memory Platform requires a shift from thinking about "chat history" to thinking about "knowledge objects."

The Four Pillars

Memories (Blocks): Discrete, atomic pieces of knowledge (facts, preferences, insights).
Subjects: The entities (Users, Projects, Organizations) that memories are linked to.
Knowledge Graph: The web of relationships between subjects and memories.
Retrieval: The dual-engine process (Vector + Graph) that finds the right context for the AI.

A Memory Block is the fundamental unit of storage. It contains:

Kind: The classification (e.g., fact, preference, insight, summary).
Content: Both natural language text (for LLMs) and structured JSON (for logic).
Scores:
- Salience: How important or "loud" this memory is.
- Stability: How likely this memory is to change (e.g., birthdays vs. current project).
- Confidence: How certain the extractor was when creating it.
Tags: Concepts or topics (e.g., #budget, #ux, #urgent) used for high-level filtering.

Every memory is about something.

Subjects have a type and an id.
Primary Dimensions are the top-level categories your platform defines (e.g., "Tenants", "Cases").
Selecting a dimension scopes the entire platform to knowledge related to that specific axis.

The platform uses a Dual-Layer Retrieval strategy:

Vector Search (Semantic)	Graph Traversal (Relational)
Finds things that sound like your query.	Finds things that are connected to your query.
"Show me preferences about travel."	"Show me everyone who works at Company X."
Handled by Qdrant.	Handled by JanusGraph.

Ingestion: Data arrives from chat, documents, or tools.
Extraction: The Curator Worker distills raw data into structured Memory Blocks.
Retrieval: The Engine Service finds relevant blocks for an AI response.
Decay: Over time, less important memories lose salience if not accessed.
Supersession: Newer, more accurate memories replace older ones via a SUPERSEDES link.