ReasonKit Memory Data Models

Version: 0.1.0

Core Concepts

ReasonKit Memory uses a hierarchical data model optimized for RAG (Retrieval-Augmented Generation) and long-term agent memory.

1. MemoryUnit (The Atom)

The fundamental unit of storage.

#![allow(unused)]
fn main() {
struct MemoryUnit {
    id: Uuid,
    content: String,
    metadata: HashMap<String, Value>,
    embedding: Vec<f32>,
    timestamp: DateTime<Utc>,
    source_uri: Option<String>,
}
}

2. Episodic Memory

Stores sequences of events or interactions.

Structure: Time-ordered list of MemoryUnits.
Use Case: Chat history, activity logs.
Indexing: Chronological + Semantic.

3. Semantic Memory

Stores facts, concepts, and generalized knowledge.

Structure: Graph-based or clustered vector space.
Use Case: “What is the capital of France?”, “User prefers dark mode”.
Indexing: RAPTOR (Recursive Abstractive Processing for Tree-Organized Retrieval).

RAPTOR Tree Structure

For large knowledge bases, we use a RAPTOR tree:

Leaf Nodes: Original chunks of text (MemoryUnit).
Parent Nodes: Summaries of child nodes.
Root Node: High-level summary of the entire cluster/document.

Retrieval traverses this tree to find the right level of abstraction for a query.

Vector Schema

Dimensions: 1536 (default, compatible with OpenAI text-embedding-3-small) or 768 (local models).
Metric: Cosine Similarity.
Engine: Qdrant / pgvector (pluggable).

Keyboard shortcuts