# Hebbian Mind Enterprise **Memory that learns. Connections that fade.** An MCP server that builds knowledge graphs through use. Concepts connect when they activate together. Unused connections decay. The more you use it, the smarter it gets. --- ## What It Does - **Associative Memory** - Save content. Query content. Related concepts surface automatically. - **Hebbian Learning** - Edges strengthen through co-activation. No manual linking required. - **Concept Nodes** - 100+ pre-defined enterprise concepts across Systems, Security, Data, Operations, and more. - **MCP Native** - Works with Claude Desktop, Claude Code, any MCP-compatible client. --- ## Installation Three paths. Pick what fits. ### Windows (Native) ```powershell # Clone the repo git clone https://github.com/For-Sunny/hebbian-mind-enterprise.git cd hebbian-mind-enterprise # Install with pip pip install -e . # Verify python -m hebbian_mind.server ``` The server runs on stdio. Press Ctrl+C to stop. ### Linux / macOS (Native) ```bash # Clone the repo git clone https://github.com/For-Sunny/hebbian-mind-enterprise.git cd hebbian-mind-enterprise # Install with pip (use a virtual environment if you prefer) pip install -e . # Verify python -m hebbian_mind.server ``` Linux gets automatic RAM disk support via `/dev/shm` when enabled. ### Docker (Teams / Enterprise) ```bash # Clone the repo git clone https://github.com/For-Sunny/hebbian-mind-enterprise.git cd hebbian-mind-enterprise # Copy environment template cp .env.example .env # Build and start docker-compose up -d # View logs docker-compose logs -f hebbian-mind ``` For RAM disk optimization: ```bash docker-compose --profile ramdisk up -d ``` --- ## Claude Desktop Integration Add to your `claude_desktop_config.json`: **Native Install:** ```json { "mcpServers": { "hebbian-mind": { "command": "python", "args": ["-m", "hebbian_mind.server"] } } } ``` **Docker Install:** ```json { "mcpServers": { "hebbian-mind": { "command": "docker", "args": ["exec", "-i", "hebbian-mind", "python", "-m", "hebbian_mind.server"] } } } ``` Restart Claude Desktop. The tools appear automatically. --- ## Configuration Environment variables control behavior. Set them before running, or use `.env` with Docker. ### Core Settings | Variable | Default | Description | |----------|---------|-------------| | `HEBBIAN_MIND_BASE_DIR` | `./hebbian_mind_data` | Data storage location | | `HEBBIAN_MIND_RAM_DISK` | `false` | Enable RAM disk for faster reads | | `HEBBIAN_MIND_RAM_DIR` | `/dev/shm/hebbian_mind` (Linux) | RAM disk path | ### Hebbian Learning | Variable | Default | Description | |----------|---------|-------------| | `HEBBIAN_MIND_THRESHOLD` | `0.3` | Activation threshold (0.0-1.0) | | `HEBBIAN_MIND_MAX_WEIGHT` | `10.0` | Maximum edge weight cap | > **Deprecated:** `HEBBIAN_MIND_EDGE_FACTOR` is no longer used. The asymptotic learning formula (LEARNING_RATE = 0.1) replaced the old harmonic strengthening factor. The env var still loads without error but has no effect on edge weights. ### Optional Integrations | Variable | Default | Description | |----------|---------|-------------| | `HEBBIAN_MIND_FAISS_ENABLED` | `false` | Enable FAISS semantic search | | `HEBBIAN_MIND_FAISS_HOST` | `localhost` | FAISS tether host | | `HEBBIAN_MIND_FAISS_PORT` | `9998` | FAISS tether port | | `HEBBIAN_MIND_PRECOG_ENABLED` | `false` | Enable PRECOG concept extraction | --- ## MCP Tools Eight tools. All available through any MCP client. ### save_to_mind Store content with automatic concept activation and edge strengthening. ```json { "content": "Microservices architecture enables independent deployment", "summary": "Optional summary", "source": "ARCHITECTURE_DOCS", "importance": 0.8 } ``` Activates matching concept nodes. Strengthens edges between co-activated concepts. ### query_mind Query memories by concept nodes. ```json { "nodes": ["architecture", "deployment"], "limit": 20 } ``` Returns memories that activated those concepts. ### analyze_content Preview which concepts would activate without saving. ```json { "content": "API authentication using JWT tokens", "threshold": 0.3 } ``` ### get_related_nodes Get concepts connected via Hebbian edges. ```json { "node": "security", "min_weight": 0.1 } ``` Returns the neighborhood graph - concepts that have fired together with "security". ### list_nodes List all concept nodes, optionally filtered. ```json { "category": "Security" } ``` ### mind_status Server health and statistics. ```json {} ``` Returns node count, edge count, memory count, strongest connections, dual-write status. ### faiss_search Semantic search via external FAISS tether (if enabled). ```json { "query": "authentication patterns", "top_k": 10 } ``` ### faiss_status Check FAISS tether connection status. --- ## Temporal Decay Memories and edges both decay over time unless reinforced. **Memory decay:** Same formula as CASCADE and PyTorch Memory. Memories lose effective importance over time. Accessed memories reset their clock. Immortal memories (importance >= 0.9) never decay. **Edge decay:** Connections between concepts weaken if not co-activated. This is the inverse of Hebbian learning -- "neurons that stop firing together, stop wiring together." Edges decay toward a minimum weight (0.1), never to zero, preserving the structure of learned associations. ### Decay Configuration | Variable | Default | Description | |----------|---------|-------------| | `HEBBIAN_MIND_DECAY_ENABLED` | `true` | Enable memory decay | | `HEBBIAN_MIND_DECAY_BASE_RATE` | `0.01` | Base exponential decay rate | | `HEBBIAN_MIND_DECAY_THRESHOLD` | `0.1` | Memories below this are hidden | | `HEBBIAN_MIND_DECAY_IMMORTAL_THRESHOLD` | `0.9` | Memories at or above this never decay | | `HEBBIAN_MIND_DECAY_SWEEP_INTERVAL` | `60` | Minutes between sweep cycles | | `HEBBIAN_MIND_EDGE_DECAY_ENABLED` | `true` | Enable edge weight decay | | `HEBBIAN_MIND_EDGE_DECAY_RATE` | `0.005` | Edge decay rate (slower than memory decay) | | `HEBBIAN_MIND_EDGE_DECAY_MIN_WEIGHT` | `0.1` | Minimum edge weight floor | Decayed memories are hidden from `query_mind` by default. Pass `include_decayed: true` to retrieve them. --- ## Architecture ### Dual-Write Pattern - **Write**: Disk first (crash-safe) -> RAM second (speed) - **Read**: RAM (instant) with disk fallback - **Startup**: Copies disk to RAM if RAM is empty Disk commits before RAM updates. If the RAM write fails, the data is already on disk -- the failure gets logged but nothing is lost. This order guarantees durability. A power loss mid-write never leaves you with RAM-only data that never reached disk. RAM disk is optional. Without it, reads and writes go directly to SQLite on disk. ### Concept Nodes 100+ pre-defined nodes across categories: - **Systems & Architecture** - service, api, component, integration - **Security** - authentication, authorization, encryption, access - **Data & Memory** - database, cache, persistence, schema - **Logic & Reasoning** - pattern, rule, validation, analysis - **Operations** - workflow, pipeline, monitoring, health - **Quality** - performance, reliability, scalability, test Nodes have keywords and prototype phrases. Content activates nodes when keywords match. ### Hebbian Learning When concepts co-activate (appear in the same saved content): 1. Edge created if none exists (initial weight: 0.15) 2. Existing edges strengthen via asymptotic formula: ``` delta = (MAX_WEIGHT - current_weight) * LEARNING_RATE new_weight = current_weight + delta ``` Each co-activation closes 10% of the gap between current weight and MAX_WEIGHT (10.0). An edge at 2.0 gains 0.8. An edge at 9.0 gains 0.1. Edges approach the ceiling but never hit it -- no saturation, no runaway weights. Combined with time-based decay (idle edges lose 2% per tick) and homeostatic scaling (total edge weight per node stays near 50.0), the graph self-regulates. Active paths strengthen. Neglected paths fade. The topology stays meaningful. "Neurons that fire together, wire together." --- ## Troubleshooting ### Server won't start Check Python version (requires 3.10+): ```bash python --version ``` Verify MCP SDK installed: ```bash pip install mcp ``` ### No activations on save Content must match node keywords above threshold. Lower the threshold: ```bash export HEBBIAN_MIND_THRESHOLD=0.2 ``` Or check what would activate: ```json {"tool": "analyze_content", "content": "your text here"} ``` ### Docker container won't connect Ensure container is running: ```bash docker ps | grep hebbian-mind ``` Check logs: ```bash docker-compose logs hebbian-mind ``` ### High memory with RAM disk Check node/edge counts via `mind_status`. Consider increasing `HEBBIAN_MIND_THRESHOLD` to activate fewer nodes, or lower `HEBBIAN_MIND_MAX_WEIGHT` to limit edge growth. --- ## Performance | Metric | Value | Notes | |--------|-------|-------| | Save latency | <10ms | Includes activation, Hebbian strengthening, and commit | | Query latency | <5ms | Node lookup + JOIN + sort | | RAM disk reads | <1ms | When `HEBBIAN_MIND_RAM_DISK=true` | | Analyze latency | <1ms | Content analysis without save | | Memory per node | ~1KB | SQLite row with keywords and phrases | | Memory per edge | ~100 bytes | SQLite row with weight and timestamps | | Startup (100 nodes) | <1 second | Schema creation + node loading + edge initialization | ### Reproducing Benchmarks A benchmark script is included to verify these claims on your hardware: ```bash python benchmarks/benchmark_performance.py ``` The script creates an isolated temp database, runs 200 iterations of each operation, and reports mean/median/P95/P99 latencies. Results are saved to `benchmarks/latest_results.json` with full system info for reproducibility. **Test conditions:** Disk-only mode (no RAM disk), WAL journal mode, 20 enterprise nodes, single-threaded. RAM disk mode will produce faster read latencies. --- ## Testing ```bash # Install dev dependencies pip install -e ".[dev]" # Run tests pytest # Run with coverage pytest --cov=hebbian_mind ``` --- ## Support - **Documentation**: [cipscorps.io/docs/hebbian-mind](https://cipscorps.io/docs/hebbian-mind) - **Email**: support@cipscorps.io - **Issues**: [GitHub Issues](https://github.com/For-Sunny/hebbian-mind-enterprise/issues) --- ## License MIT License. See [LICENSE](./LICENSE) for terms. --- *Memory that learns. Concepts that connect. The more you use it, the smarter it gets.* --- **Made by [CIPS Corp](https://cipscorps.io)** [Website](https://cipscorps.io) | [Store](https://store.cipscorps.io) | [GitHub](https://github.com/For-Sunny) | glass@cipscorps.io Enterprise cognitive infrastructure for AI systems: [PyTorch Memory, Soul Matrix, CMM, and the full CIPS Stack](https://store.cipscorps.io). Copyright (c) 2025-2026 C.I.P.S. LLC