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.

- 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.

# 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.

# 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/shmwhen enabled.

# 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

Add to yourclaude_desktop_config.json:

{ "mcpServers": { "hebbian-mind": { "command": "python", "args": ["-m", "hebbian_mind.server"] } } }

{ "mcpServers": { "hebbian-mind": { "command": "docker", "args": ["exec", "-i", "hebbian-mind", "python", "-m", "hebbian_mind.server"] } } }

Restart Claude Desktop. The tools appear automatically.

Environment variables control behavior. Set them before running, or use.envwith Docker.

Deprecated:HEBBIAN_MIND_EDGE_FACTORis 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.

Eight tools. All available through any MCP client.

Store content with automatic concept activation and edge strengthening.

{ "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.

{ "nodes": ["architecture", "deployment"], "limit": 20 }

Returns memories that activated those concepts.

Preview which concepts would activate without saving.

{ "content": "API authentication using JWT tokens", "threshold": 0.3 }

Get concepts connected via Hebbian edges.

{ "node": "security", "min_weight": 0.1 }

Returns the neighborhood graph - concepts that have fired together with "security".

List all concept nodes, optionally filtered.

{ "category": "Security" }

Returns node count, edge count, memory count, strongest connections, dual-write status.

Semantic search via external FAISS tether (if enabled).

{ "query": "authentication patterns", "top_k": 10 }

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.

Decayed memories are hidden fromquery_mindby default. Passinclude_decayed: trueto retrieve them.

- 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.

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.

When concepts co-activate (appear in the same saved content):
- Edge created if none exists (initial weight: 0.15)
- Existing edges strengthen via asymptotic formula:

delta = (MAX_WEIGHT - current_weight)  LEARNING_RATE new_weight = current_weight + delta

{"tool": "analyze_content", "content": "your text here"}

python benchmarks/benchmark_performance.py

# Install dev dependencies pip install -e ".[dev]" # Run tests pytest # Run with coverage pytest --cov=hebbian_mind

Website|Store|GitHub|glass@cipscorps.io

Enterprise cognitive infrastructure for AI systems:PyTorch Memory, Soul Matrix, CMM, and the full CIPS Stack.

Structured business ontology giving AI agents deterministic access to institutional reasoning — 18 MCP tools, Dolt backend, 0% → 100% "why?" recall vs Markdown+RAG.