STATUS: ACTIVE DEVELOPMENT // PROTOTYPE 8-MODE UNIVERSAL INFERENCE MATRIX 23-DIRECTORY CANONICAL MEMORY VAULT

GOLEM — Graph-Optimized Execution Mind. Local-First Memory. Every Way to Run AI.

GOLEM is Built by Kellz's flagship local-first and hybrid AI platform architecture. It pairs a plain-text, human-inspectable 23-directory Markdown Memory Vault with an 8-Mode Universal Inference Matrix—allowing the exact same knowledge graph, skills, and safety guardrails to run on a thin-client laptop (0 MB local model RAM) or a multi-GPU workstation running llama.cpp, Ollama, ONNX DirectML, or vLLM / SGLang.

Official GOLEM Gothic Cyber Luxury G-Sigil
Official GOLEM Sigil
OBSIDIAN • CHAMPAGNE GOLD • CRIMSON

Distinct product mark representing GOLEM under the Built by Kellz systems ecosystem.

[ HARDWARE-HONEST RUNTIME // 8 EXECUTION MODES ]

Universal Inference Matrix: From Thin Client to ML Workstation

Most AI desktop apps either force a heavy local download that crashes an 8 GB laptop, or lock ML engineers out of low-level sampler and KV-cache controls. GOLEM adapts to your hardware profile automatically.

MODE 0 // THIN-CLIENT DEFAULT 0 MB LOCAL MODEL RAM

API-Only / Thin-Client Mode (Low-Memory Laptops & Phones)

For devices with <8 GB RAM or integrated graphics, GOLEM disables local weight loading entirely—keeping memory footprint under ~150 MB while providing full access to the 23-folder Memory Vault, graph visualization, and BYOK cloud/LAN endpoints (Gemini, OpenAI, Anthropic, Groq, OpenRouter, or a home server over Tailscale).

MODE 1 // NATIVE GGUF CUDA • VULKAN • METAL • CPU

Managed llama.cpp (llama-server + llama-swap)

Direct binary process supervision of llama-server with automatic GGUF binary header inspection (reading exact block counts, context length, and KV-cache requirements before allocation) and zero-downtime model switching via llama-swap.

MODE 2 // ZERO-CONFIG DAEMON PORT 11434

Native Ollama Daemon Adapter

Auto-discovers local or LAN Ollama instances, enumerates installed tags and running VRAM residency via /api/ps, and controls keep_alive eviction so models unload cleanly when idle.

MODE 3 // DESKTOP ECOSYSTEM PORTS 1234 • 1337 • 5001

LM Studio, Jan, LocalAI & KoboldCpp Bridge

Connects to any OpenAI-compatible local server you already run—eliminating duplicate model downloads by sharing weights across your existing local AI tools.

MODE 4 // WINDOWS GPU & NPU DIRECTML • QNN • VINO

ONNX Runtime GenAI & DirectML / NPU Acceleration

Hardware-native execution on Windows Copilot+ NPUs, AMD Radeon, Intel Arc, and NVIDIA GPUs via DirectML—ideal for low-wattage background classification, embeddings, and voice transcription.

MODE 5 // BURST GPU TUNNEL 30 GB VRAM BURST

Remote Kaggle Dual-T4 & Colab GPU Bridge

Authenticated SSH/tunnel bridge to remote notebook GPUs (such as Dual NVIDIA T4s with 30 GB combined VRAM) for heavy batch indexing or 30B–70B model runs without local GPU wear.

MODE 6 // ML ENGINEER GRADE PAGEDATTENTION • RADIXCACHE

vLLM, SGLang, TensorRT-LLM & ExLlamaV2

Exposes deep inference controls for ML engineers: Guided Decoding (JSON Schema, Regex, EBNF grammar constraints), dynamic LoRA adapter hot-swapping, speculative decoding telemetry, and token-level logprob inspection.

MODE 7 // MULTI-MODEL ARENA SUBPROCESS & DAG EVAL

Custom CLI Subprocess & Multi-Model Graph Arena

Bind any custom Python/Rust/C++ inference binary via stdin/stdout or HTTP, and dispatch the same prompt + Memory Vault context across multiple runtimes side-by-side to compare TTFT, tokens/sec, and structural fidelity.

[ STRUCTURED KNOWLEDGE // 23-DIRECTORY VAULT ]

Plain-Text Markdown Memory That Outlives Any Model

Models change every quarter; your institutional memory should never be trapped inside an opaque SaaS vector database. GOLEM stores every rule, architecture decision, agent definition, checkpoint, and incident report in a strictly numbered, Obsidian-compatible Markdown Memory Vault.

Every file is UTF-8 Markdown with explicit [[Wikilinks]], YAML frontmatter, and deterministic folder routing so both human operators and local AI agents can audit the exact source of truth.

[ Commission a Structured Knowledge Vault for Your Team → ]
CANONICAL_MEMORY_VAULT // 00–22 + 99 UTF-8 MARKDOWN
00_SYSTEM — Operating Rules & Index
01_VISION — North Star & Doctrine
02_ARCHITECTURE — Core Specs
03_RUNTIME — Execution Matrices
04_MEMORY — Episodic & Semantic
05_PROJECTS — Active Workspaces
06_AGENTS — Bounded Personas
07_SKILLS — Reusable Workflows
08_TOOLS — Deterministic Scripts
09_MCP — Model Context Servers
10_KNOWLEDGE — Domain Reference
12_GRAPH — Node & Edge Topologies
14_CORTEX_RT — Real-Time Loops
19_INCIDENTS — Failure Postmortems
21_V2 — V2 Cockpit & Inference Specs
99_ARCHIVE — Immutable Cold Storage
[ CUSTOM LOCAL & HYBRID AI CONSULTING ]

Want Private, Grounded AI Infrastructure for Your Organization?

While the standalone GOLEM desktop application remains in active prototype development, Kelly Greene designs and implements custom local/hybrid inference setups, structured Markdown knowledge vaults, and bounded RAG assistants for businesses today.