ReasonKit Mem

Memory & Retrieval Infrastructure for ReasonKit

The Long-Term Memory Layer ("Hippocampus") for AI Reasoning

Documentation | ReasonKit Core | Website

---

ReasonKit Mem is the memory layer ("Hippocampus") for ReasonKit. It provides vector storage, hybrid search, RAPTOR trees, and embedding support.

Features

• Vector Storage - Qdrant-based dense vector storage with embedded mode
• Hybrid Search - Dense (Qdrant) + Sparse (Tantivy BM25) fusion
• RAPTOR Trees - Hierarchical retrieval for long-form QA
• Embeddings - Local (BGE-M3) and remote (OpenAI) embedding support
• Reranking - Cross-encoder reranking for precision

Installation

Universal Installer (Recommended)

Installs all 4 ReasonKit projects together:

curl -fsSL https://get.reasonkit.sh | bash -s -- --with-memory

Platform & Shell Support:

• ✅ All platforms (Linux/macOS/Windows/WSL)
• ✅ All shells (Bash/Zsh/Fish/Nu/PowerShell/Elvish)
• ✅ Auto-detects shell and configures PATH
• ✅ Beautiful progress visualization

Cargo (Rust Library)

Add to your Cargo.toml:

[dependencies]
reasonkit-mem = "0.1"
tokio = { version = "1", features = ["full"] }

Usage

Basic Usage (Embedded Mode)

use reasonkit_mem::storage::Storage;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Create embedded storage (automatic file storage fallback)
    let storage = Storage::new_embedded().await?;

    // Use storage...
    Ok(())
}

Storage with Custom Configuration

use reasonkit_mem::storage::{Storage, EmbeddedStorageConfig};
use std::path::PathBuf;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Create storage with custom file path
    let config = EmbeddedStorageConfig::file_only(PathBuf::from("./data"));
    let storage = Storage::new_embedded_with_config(config).await?;

    // Or use Qdrant (requires running server)
    let qdrant_config = EmbeddedStorageConfig::with_qdrant(
        "http://localhost:6333",
        "my_collection",
        1536,
    );
    let qdrant_storage = Storage::new_embedded_with_config(qdrant_config).await?;

    Ok(())
}

Hybrid Search with KnowledgeBase

use reasonkit_mem::retrieval::KnowledgeBase;
use reasonkit_mem::{Document, DocumentType, Source, SourceType};
use chrono::Utc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Create in-memory knowledge base
    let kb = KnowledgeBase::in_memory()?;

    // Create a document
    let source = Source {
        source_type: SourceType::Local,
        url: None,
        path: Some("notes.md".to_string()),
        arxiv_id: None,
        github_repo: None,
        retrieved_at: Utc::now(),
        version: None,
    };

    let doc = Document::new(DocumentType::Note, source)
        .with_content("Machine learning is a subset of artificial intelligence.".to_string());

    // Add document to knowledge base
    kb.add(&doc).await?;

    // Search using sparse retrieval (BM25)
    let results = kb.retriever().search_sparse("machine learning", 5).await?;

    for result in results {
        println!("Score: {:.3}, Text: {}", result.score, result.text);
    }

    Ok(())
}

Using Embeddings

use reasonkit_mem::embedding::{EmbeddingConfig, EmbeddingPipeline, OpenAIEmbedding};
use reasonkit_mem::retrieval::KnowledgeBase;
use std::sync::Arc;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Create OpenAI embedding provider (requires OPENAI_API_KEY env var)
    let embedding_provider = OpenAIEmbedding::openai()?;
    let pipeline = Arc::new(EmbeddingPipeline::new(Arc::new(embedding_provider)));

    // Create knowledge base with embedding support
    let kb = KnowledgeBase::in_memory()?
        .with_embedding_pipeline(pipeline);

    // Now hybrid search will use both dense (vector) and sparse (BM25)
    // let results = kb.query("semantic search query", 10).await?;

    Ok(())
}

Embedded Mode Documentation

For detailed information about embedded mode, see docs/EMBEDDED_MODE_GUIDE.md.

Architecture

The RAPTOR Algorithm (Hierarchical Indexing)

ReasonKit Mem implements RAPTOR (Recursive Abstractive Processing for Tree-Organized Retrieval) to answer high-level questions across large document sets.

The Memory Dashboard

Integration Ecosystem

Technology Stack

Component	Technology	Purpose
Qdrant	qdrant-client 1.10+	Dense vector storage
Tantivy	tantivy 0.22+	BM25 sparse search
RAPTOR	Custom Rust	Hierarchical retrieval
Embeddings	BGE-M3 / OpenAI	Dense representations
Reranking	Cross-encoder	Final precision boost

Project Structure

reasonkit-mem/
├── src/
│   ├── storage/      # Qdrant vector + file-based storage
│   ├── embedding/    # Dense vector embeddings
│   ├── retrieval/    # Hybrid search, fusion, reranking
│   ├── raptor/       # RAPTOR hierarchical tree structure
│   ├── indexing/     # BM25/Tantivy sparse indexing
│   └── rag/          # RAG pipeline orchestration
├── benches/          # Performance benchmarks
├── examples/         # Usage examples
├── docs/             # Additional documentation
└── Cargo.toml

Feature Flags

Feature	Description
default	Core functionality
python	Python bindings via PyO3
local-embeddings	Local BGE-M3 embeddings via ONNX Runtime

API Reference

Core Types (re-exported at crate root)

use reasonkit_mem::{
    // Documents
    Document, DocumentType, DocumentContent,
    // Chunks
    Chunk, EmbeddingIds,
    // Sources
    Source, SourceType,
    // Metadata
    Metadata, Author,
    // Search
    SearchResult, MatchSource, RetrievalConfig,
    // Processing
    ProcessingStatus, ProcessingState, ContentFormat,
    // Errors
    MemError, MemResult,
};

Storage Module

use reasonkit_mem::storage::{
    Storage,
    EmbeddedStorageConfig,
    StorageBackend,
    InMemoryStorage,
    FileStorage,
    QdrantStorage,
    AccessContext,
    AccessLevel,
};

Embedding Module

use reasonkit_mem::embedding::{
    EmbeddingProvider,      // Trait for embedding backends
    OpenAIEmbedding,        // OpenAI API embeddings
    EmbeddingConfig,        // Configuration
    EmbeddingPipeline,      // Batch processing pipeline
    EmbeddingResult,        // Single embedding result
    EmbeddingVector,        // Vec<f32> alias
    cosine_similarity,      // Utility function
    normalize_vector,       // Utility function
};

Retrieval Module

use reasonkit_mem::retrieval::{
    HybridRetriever,        // Main retrieval engine
    KnowledgeBase,          // High-level API
    HybridResult,           // Search result
    RetrievalStats,         // Statistics
    // Fusion
    FusionEngine,
    FusionStrategy,
    // Reranking
    Reranker,
    RerankerConfig,
};

Version & Maturity

Component	Status	Notes
Vector Storage	✅ Stable	Qdrant integration production-ready
Hybrid Search	✅ Stable	Dense + Sparse fusion working
RAPTOR Trees	✅ Stable	Hierarchical retrieval implemented
Embeddings	✅ Stable	OpenAI API fully supported
Local Embeddings	🔶 Beta	BGE-M3 ONNX (enable with local-embeddings feature)
Python Bindings	🔶 Beta	Build from source with --features python

Current Version: v0.1.2 | CHANGELOG | Releases

Verify Installation

use reasonkit_mem::storage::Storage;

#[tokio::main]
async fn main() -> anyhow::Result<()> {
    // Quick verification - creates in-memory storage
    let storage = Storage::new_embedded().await?;
    println!("ReasonKit Mem initialized successfully!");
    Ok(())
}

License

Apache License 2.0 - see LICENSE

---

Part of the ReasonKit Ecosystem

ReasonKit Core | ReasonKit Web | Website

"See How Your AI Thinks"