Skip to main content

What is Memory?

In the Memory Platform, a Memory is a discrete, retrievable piece of knowledge that an AI can recall to provide context, personalization, and continuity.

Unlike chat history, which is a raw transcript of messages, memories are synthesized knowledge extracted from various sources.

Key Characteristics​

  • Discrete: Each memory represents a single fact, preference, insight, or summary.
  • Structured: Beyond text, memories can contain structured data for programmatic use.
  • Contextual: Every memory is associated with a subject (Who is this about?) and a source (Where did this come from?).
  • Dynamic: Memories have scores (salience, confidence, stability) that evolve over time.
  • Linked: Memories form a knowledge graph, connecting entities and concepts.

Why AI Needs Memory​

  1. Personalization: Remember user preferences and history across sessions.
  2. Knowledge Accumulation: Build a long-term understanding of projects, organizations, and topics.
  3. Context Continuity: Avoid repeating questions or losing track of decisions.
  4. Efficiency: Retrieve only relevant information instead of processing massive histories.

The Mental Model​

Think of the Memory Platform as a Knowledge Middleware between your data sources and your AI:

graph LR
    S[Sources: Chat, Docs, Tools] --> E[Extraction Layer]
    E --> M[Memory Store]
    M --> R[Retrieval Layer]
    R --> A[AI Response]
  1. Extraction: Raw data is processed to create structured memories.
  2. Storage: Memories are stored in both a Knowledge Graph (for relationships) and a Vector Store (for semantic search).
  3. Retrieval: When an AI needs context, the system finds the most relevant and salient memories.
  4. Usage: Memories are injected into the AI's prompt to guide its behavior and answers.