⚡ Spector — The AI Memory Backbone¶

Agent-ready cognitive memory that forms associations — sub-millisecond recall, zero infrastructure.

Spector gives AI agents real memory: it remembers, forgets, consolidates, and forms associations across working, episodic, semantic, and procedural tiers, linked by Hebbian, temporal, and entity graphs. Retrieval fuses dense semantic search with hybrid signals and cognitive scoring for sub-millisecond recall.

Connect your agents through the built-in MCP server (Claude Desktop, Cursor, custom agents), call it over REST/gRPC, use the Python SDK, or embed it as a single JAR — no external database, no infrastructure to run. Every user, agent, or tenant is physically isolated in its own on-disk namespace. Java Project Panama keeps it all off-heap with zero GC pressure.

🔥 Key Numbers¶

🗺️ Choose Your Path¶

💡 How It Works¶

Spector fuses semantic vector search, hybrid retrieval, and cognitive scoring into a single pipeline:

graph LR
    A["🤖 AI Agent"] --> B["📡 MCP Server"]
    B --> C["⚡ SpectorEngine"]
    C --> D["🧠 Hybrid Search"]
    D --> E["🎯 RRF Fusion"]
    E --> F["🤖 LLM Re-ranking"]
    F --> G["✨ Results"]

    H["📄 Document"] --> I["🧩 Chunking"]
    I --> J["🧬 Embedding"]
    J --> C

What Makes Spector Different¶

• Flexible deployment — connect over MCP or REST/gRPC, drive it from the Python SDK, or embed it as a library inside your JVM. No Docker, no external database, no network hops when embedded.
• Agent-native — 16 MCP tools for memory, recall, and cognitive operations. Connect Claude Desktop or Cursor in one config line.
• Associative memory — Hebbian co-activation, temporal chains, and entity graphs with spreading activation, so recall surfaces what's related, not just what matches.
• Cognitive memory — the only system combining power-law decay, Two-Factor strengthening (Bjork & Bjork), emotional valence, and Hebbian association in a single scoring formula.
• Zero GC pressure — all vector data and headers live off-heap via Project Panama. The JVM garbage collector never sees memory records.
• SIMD everywhere — vector distance, quantization, and scoring use Java Vector API (AVX2/AVX-512/NEON) for hardware-accelerated computation.

🌟 Project Stats¶

Built with ⚡ by Spectrayan · GitHub · Apache 2.0 · BSL 1.1 (memory)