🌟 What is Spector?¶

The Zero-Overhead, Agent-Ready AI Memory Backbone.

Legacy AI stacks bolt memory onto stateless vector databases — storage without cognition. Spector is built from the ground up for modern AI agents: it remembers, forgets, consolidates, and forms associations across a biologically-inspired memory graph — Hebbian co-activation, temporal chains, and entity links — then retrieves with fused semantic and hybrid scoring. Connect any AI agent through the built-in MCP server, call it over REST/gRPC, drive it from the Python SDK, or embed it directly in the JVM.

Spector is an open-source, high-performance cognitive memory system. It delivers sub-millisecond memory retrieval, native AI agent integration, and zero infrastructure complexity — reach it from any language over MCP or REST/gRPC, use the Python SDK, or embed it as a single JAR. Every user, agent, or tenant is physically isolated in its own on-disk namespace. Under the hood, modern Java 25, Project Panama, and the Vector API deliver the performance.

🎯 What It Does¶

Spector indexes documents with their vector embeddings and text content, then retrieves them using multiple strategies — directly from AI agents or your application code:

graph LR
    subgraph Clients
        MCP["🤖 AI Agent (MCP)"]
        REST["🌐 REST API"]
        SDK["📦 Java SDK"]
    end

    subgraph Recall Modes
        HYBRID[Hybrid Recall] --> D[RRF Fusion]
        VECTOR[Dense Recall] --> E[HNSW ANN]
        KEYWORD[Lexical Recall] --> F[BM25 Scoring]
        SPLADE[Sparse Recall] --> G[Splade Index]
        COLBERT[ColBERT Rerank] --> H[MaxSim Rerank]
    end

    D & E & F & G & H --> Results[Results]

    MCP & REST & SDK --> HYBRID & VECTOR & KEYWORD & SPLADE & COLBERT

Mode	Active Layers	How It Works	Best For
🧠 Dense Recall	Dense Vector (HNSW)	HNSW approximate nearest neighbor graphs	Semantic similarity, conceptual queries
📝 Lexical Recall	BM25 only	SIMD-accelerated Lucene-style scoring	Exact term matching, error codes, IDs
🧬 Hybrid Recall	BM25 + Dense Vector	Parallel execution fused via RRF	General purpose, balanced quality
📈 Sparse Recall	SPLADE / Li-LSR	Neural term expansion with inverted indexing	Synonym-aware lexical search
🚀 ColBERT Rerank	Vector + BM25 + ColBERT	Late-interaction MaxSim token matching	Maximum precision, grounded context
🏛️ SpectorIndex	IVF-HNSW-SVASQ	Adaptive quantized hybrid index	Large scale index compression + recall

💎 Key Differentiators¶

🤖 Agent-Native (MCP Protocol)¶

Includes a built-in Model Context Protocol server with 16 cognitive memory tools. AI agents connect directly via JSON-RPC — no adapter layer, no network round-trips.

Feature	Python Vector DB MCP	Spector MCP
Recall latency	2–10ms	Ultra-low latency (in-process SIMD) †
Network overhead	HTTP/gRPC round-trip	Zero (in-process)
Concurrent queries	Limited by Python GIL	61,000 QPS †
Dependencies	Python framework stack	Single JAR

† Measured. See Benchmarks.

Tip

See the MCP Server Guide to connect Claude Desktop, Cursor, or any MCP client in minutes.

� Associative Cognitive Graphs¶

Spector doesn't just store vectors — it links memories. Hebbian co-activation, temporal chains, and an LLM-powered entity graph connect related memories, and spreading activation means recall surfaces what's related, not just what matches. It's memory that forms associations, the way a brain does.

🔒 Physical Namespace Isolation¶

Every user, agent, or tenant's memory lives in its own on-disk directory tree — true data separation, not a WHERE tenant_id = ? filter over a shared store. Namespaces are hash-sharded to scale to millions and encrypted at rest (AES-256-GCM).

�📦 Pure-Java Engine, Zero Dependencies¶

Unlike most vector databases that rely on C++, Rust, or Python bindings, Spector's engine is pure Java — no JNI, no native libraries, no external infrastructure to install. It uses the JDK's own Vector API for SIMD acceleration.

Tip

Add the JAR to your classpath and you're done. No Docker, no clusters, no ops.

🚀 Modern JVM Technologies¶

Technology	Purpose
Java Vector API	SIMD-accelerated math (AVX2/AVX-512/NEON)
Panama FFM	Zero-copy memory-mapped storage, GPU interop
Virtual Threads	Millions of concurrent operations without thread pools
Structured Concurrency	Safe parallel task management

⚡ Sub-Millisecond at Scale¶

HNSW at 100K documents (128 dimensions, top-10, M=16, efSearch=64):

Recall Type	Average Latency	Throughput
Dense	Ultra-fast	High QPS
Lexical	0.98 ms	1,019 QPS
Hybrid	1.01 ms	994 QPS

SpectorIndex (IVF-HNSW-SVASQ) at 10K documents (4096-dim real Qwen3 embeddings):

Config	Average Latency	Throughput	Recall@10
nCentroids=128, nProbe=4	0.46 ms	2,173 QPS	1.0000
nCentroids=64, nProbe=4	0.62 ms	1,601 QPS	1.0000
nCentroids=128, nProbe=16	1.26 ms	792 QPS	1.0000

Note

SpectorIndex achieves perfect recall while searching only 3.1% of the data (nProbe=4 out of 128 centroids). Ingestion is 28–160× faster than standalone HNSW. Numbers measured on 24-core x86, AVX2, Java 25, ZGC with Qwen3-embedding real vectors. For comprehensive, multi-centroid sweeps and adaptive HNSW shard promotion benchmarks, see the dedicated Large-Scale Real-Embedding Benchmarks page.

🏠 Dual Deployment Modes¶

Mode	Description	Best For
Embedded	In-process library, zero network overhead	Microservices, desktop apps, edge
Server	REST API with CORS, auth, and metrics	Teams, multi-language clients

🗜️ Advanced Quantization (SVASQ + IVF-PQ)¶

Spector offers two quantization paths:

SVASQ (Vectorized Affine Scalar Quantization): Uses the Fast Walsh-Hadamard Transform to spread variance before INT8 quantization, achieving 4× compression with near-lossless recall (~97–99.5%). Used inside SpectorIndex shards.
IVF-PQ (Product Quantization): Provides 32× memory compression for billion-scale datasets.

Important

SVASQ gives INT8 the precision of INT12–16 by rotating vectors before quantization. See the SVASQ Deep Dive for the full theory.

📊 How Spector Compares¶

Latency Comparison (100K docs, 128-dim, top-10)¶

Engine	Language	Vector Avg	Vector P99
⚡ Spector	Java 25	Ultra-low	Sub-millisecond
hnswlib	C++	0.1–0.5 ms	~1 ms
FAISS	C++	0.2–0.8 ms	1–2 ms
Lucene 9+	Java	1–5 ms	5–10 ms
Elasticsearch 8+	Java	2–10 ms	10–25 ms
Qdrant	Rust	2–5 ms	10–25 ms
Milvus	Go/C++	3–10 ms	10–35 ms

Note

Spector's vector recall latency is competitive with native C++ implementations (hnswlib, FAISS) for in-process workloads. Numbers for external systems are from published benchmarks and ann-benchmarks.com. Hardware and configuration differences apply — these are directional comparisons, not controlled A/B tests.

Feature Comparison¶

Feature	Spector	Elasticsearch	Qdrant	Milvus	hnswlib
Deployment	Embedded + Server	Cluster only	Server only	Cluster only	Embedded only
MCP Server	✅ Built-in (16 tools)	❌	❌	❌	❌
Hybrid Recall	✅ RRF built-in	✅ RRF	✅ Sparse+Dense	✅ RRF	❌
Zero Dependencies	✅ JDK only	❌ Heavy stack	❌ Tokio runtime	❌ etcd, MinIO, Pulsar	✅ Header-only
Virtual Threads	✅ Project Loom	❌ Platform threads	N/A (Rust async)	N/A (Go goroutines)	N/A
GPU Acceleration	✅ CUDA (Panama FFM)	❌	✅ Vulkan (indexing)	✅ CUDA (search + indexing)	❌
Quantization	✅ Scalar INT8 + IVF-PQ	✅ BBQ + Scalar + DiskBBQ (IVF)	✅ Scalar + Binary	✅ IVF-PQ + IVF-SQ	❌
Re-ranking	✅ ColBERT v2 (FFM SIMD)	✅ Elastic Rerank + Inference API	✅ FastEmbed / ColBERT	✅ vLLM Ranker + Cross-encoder	❌
Distributed	✅ gRPC fan-out	✅ Built-in sharding	✅ Raft consensus	✅ gRPC + etcd	❌
SIMD Acceleration	✅ Java Vector API	✅ simdvec (Panama)	✅ Native SIMD	✅ AVX/NEON	✅ AVX/SSE

Note

This comparison reflects publicly available information as of May 2025. Feature availability may vary by version and deployment mode. All products are actively evolving.

🛠️ Use Cases¶

🤖 Agentic AI Memory¶

Connect AI agents (Claude, Cursor, custom) directly to Spector via the built-in MCP server. The agent autonomously ingests documents, searches for relevant context, and retrieves information — all with zero glue-code. "Point your LLM at Spector's MCP port, and it instantly has mathematically-perfect long-term memory."

🔍 Semantic Search & Recall Applications¶

Power product search, documentation recall, code search, or any application where meaning matters more than exact keywords.

💡 Recommendation Systems¶

Use similarity to find items similar to what users have engaged with. Sub-millisecond latency makes real-time recommendations practical.

🏢 Hybrid Enterprise Memory¶

Combine keyword precision (finding exact product SKUs, error codes) with semantic understanding (finding conceptually related documents).

📱 Embedded Analytics¶

Drop Spector into existing Java applications without infrastructure changes. Perfect for desktop applications, microservices, or edge deployments.

✅ When to Choose Spector¶

Note

Choose Spector when:

You want AI agents to autonomously manage their memories (MCP integration)
You want sub-millisecond hybrid recall without infrastructure complexity
You work in any language — connect over MCP or REST/gRPC, drive it from the Python SDK, or embed it natively in the JVM
You need cognitive memory that forms associations, not just a vector store
You want GPU acceleration without leaving the JVM
Zero external dependencies matters to your deployment

Warning

Consider alternatives when:

You need a managed cloud service with zero ops
You need multi-modal retrieval across images, audio, and video out of the box
You need built-in ML model serving

🚀 Next Steps¶

Getting Started — Build and run your first recall in 5 minutes
MCP Server Guide — Connect an AI agent in 3 steps
Architecture Overview — Understand how it works under the hood
REST API Reference — Full API documentation
Core Concepts — Deep dive into the algorithms