RefAlfred - Portable OpenClaw Fork with Associative Memory

Alfred - Portable OpenClaw Fork with Associative Memory

Alfred - Portable OpenClaw Fork with Associative Memory

AI-first business OS for solopreneurs. Now fully portable: deploys to Hetzner VM in 10 minutes, self-fixes bugs, maintains an associative memory vault with auto-clustering.

What Is It?

Alfred is a modded OpenClaw fork that's evolved into "AlfredOS" - an AI-first business operating system for solopreneurs. It combines:

  1. Full Infrastructure Deployment - Auto-deploys entire stack to Hetzner VM
  2. Associative Memory Vault - Structured knowledge base with auto-clustering
  3. Self-Fixing Deployment - Alfred fixes bugs automatically during deployment
  4. Business App Integration - LibreChat, n8n, Ghost, Cal.com, NocoDB

Key Milestone (Feb 5, 2026): Alfred is now portable. The entire modded OpenClaw fork can deploy from scratch in ~10 minutes.

"The whole setup takes about 10 minutes. In case of any bugs, Alfred fix the deployment automatically and pushes it to completion while I sip my coffee."

What Problem Does It Solve?

Solopreneur Infrastructure Pain

  • SaaS fees - Constant subscription costs
  • Server management - DevOps complexity
  • Context switching - Copy/paste between tools
  • Memory fragmentation - Information scattered across tools

Alfred's Solution

  • One-click deployment - Full stack in minutes
  • Self-hosted - Own your data, pay token costs only
  • Integrated AI agent - Access all tools natively
  • Associative memory - Unified knowledge vault with auto-maintenance

How Alfred's Portable Deployment Works

1. Infrastructure Deployment

  • Deploys to: Hetzner VM
  • Secrets management: Pulumi ESC
  • Networking: Tailscale VPN
  • Time: ~10 minutes

2. Vault Preloading

  • Loads custom agents
  • Installs skills and scripts
  • Sets up vault maintenance automation

3. Autoclustering Pipeline

  • Embedding: nomic-embed-text
  • Clustering: HDBSCAN on vector data to reveal patterns
  • Labeling: Cerebras using Llama 3.1 for cluster and relationship labels
  • Ontology: Palantir-based structure for nodes

4. Self-Fixing

Alfred monitors deployment, detects bugs, fixes them automatically, and pushes to completion.

The Associative Memory System

Alfred builds and maintains an associative memory vault - a structured knowledge base that:

Automatic Processing

  • Processes conversations and files at regular intervals into the vault
  • Follows ontology strictly - Palantir-based structure
  • Accepts any input - Audio, video, text braindumps

Automatic Maintenance

  • Scans for violations - Ontology issues, duplicates, empty stubs, broken links
  • Fixes automatically - No manual intervention needed
  • Finds hidden relationships - Links unlinked entities
  • Enriches content - Using web search

Clustering Workflow

  1. Embed full vault text with nomic-embed-text
  2. Run HDBSCAN on vector data to reveal clusters
  3. Label clusters with Qwen 2.5 7B (running locally)
  4. Label relationships and relationship types (also Qwen)
  5. Reruns regularly on DGX Spark to keep vault healthy

Fast Retrieval with Small Language Models

Alfred trains small language models (Qwen 3 0.6B) on the ontology schema (not content) for sub-100ms retrieval:

The Process:

  1. Use Opus 4.5 and Sonnet 4.5 subagents to generate 1000 pairs of:
    • Natural language prompts
    • Obsidian Base syntax responses
  2. Submit TRL job with LoRA fine-tuning on HF Jobs (a10g-large)
  3. Convert to Q4_K_M quantization for Ollama
  4. Deploy on DGX Spark
  5. Add alias on OpenClaw for future use

How It Works at Runtime:

  • Every prompt gets intercepted
  • Small model returns relevant Obsidian filenames in Base syntax (<100ms)
  • Filenames get injected into prompt
  • Alfred agent receives context-aware prompt
  • Result: Associative memory - rebuilding context on the fly faster than lightning

Expected accuracy: ~70%+ (theory)

The Tech Stack

Core Infrastructure

  • Agent Interface: LibreChat (custom fork called "Alfred")
  • Automation: n8n
  • Blog: Ghost
  • Calendar: Cal.com
  • Database: NocoDB
  • Deployment: Railway.com template (moving to Hetzner)
  • MCP Connections: Hardwired between agent and apps

AI/ML Stack

  • Models: Opus 4.5, Sonnet 4.5, Qwen 2.5 7B, Qwen 3 0.6B
  • Embedding: nomic-embed-text
  • Clustering: HDBSCAN
  • Labeling: Cerebras (Llama 3.1)
  • Hardware: DGX Spark

Infrastructure-as-Code

  • Secrets: Pulumi ESC
  • VPN: Tailscale
  • Hosting: Hetzner VM

Why OpenClaw?

Alfred is built as a modded OpenClaw fork because OpenClaw provides:

  • Agent coordination framework
  • MCP integrations
  • Skill system
  • Multi-agent orchestration

David extended it with:

  • Portable deployment automation
  • Associative memory vault
  • Auto-clustering pipeline
  • Self-fixing capabilities

Philosophy: AI-First Solopreneurship

Core Principles

  1. Own your data - Self-hosted, not SaaS
  2. Pay for tokens, not seats - Token costs dropped 280x in 3 years
  3. Zero manual context switching - AI has native access to all tools
  4. Write SOPs once - AI executes them repeatedly
  5. Minimize maintenance - Auto-fixing, auto-clustering, auto-enrichment

The Vision

"My goal is to run my business autonomously and AlfredOS provides the scaffolding."

Alfred represents a shift from "AI as tool" to "AI as business operating system."

The Journey (Timeline)

Early 2025: Started building "my own butler" - simple Telegram bot
Mid 2025: Realized context management needed structured data → AlfredOS v1.0
May 2025: AlfredOS v1.0 launched (Linux, Docker-in-Docker, unusable)
Aug 2025: AlfredOS v1.1 rebuilt on Railway.com, made open source
Jan 2026: Alfred gains associative memory with small language model training
Feb 2026: Alfred becomes portable - full deployment automation

Key Innovation: Self-Fixing Deployment

"In case of any bugs, Alfred fix the deployment automatically and pushes it to completion while I sip my coffee."

This is a major milestone: an AI system that can:

  1. Deploy its own infrastructure
  2. Detect deployment failures
  3. Fix them automatically
  4. Push to completion without human intervention

This moves beyond "AI agents that write code" to "AI agents that deploy and maintain themselves."

Comparison to Other AI Systems

FeatureAlfredStandard AIOpenClaw Base
DeploymentSelf-deploying (10min)ManualManual
MemoryAssociative vaultSession-basedContext layer
MaintenanceSelf-fixingManualManual
IntegrationHardwired MCPAPI callsMCP support
InfrastructureFull stackCloud APIFramework only
ClusteringAuto (HDBSCAN)NoneNone
Retrieval<100ms (SLM)RAG (~seconds)Standard

Use Cases

For Solopreneurs

  • Run entire business on self-hosted infrastructure
  • Pay token costs instead of SaaS fees
  • Own all data
  • Automate repetitive tasks with SOPs

For Builders

  • Template for AI-first business infrastructure
  • Reference implementation of associative memory
  • Example of self-fixing deployment
  • OpenClaw fork with production features

For Researchers

  • Associative memory implementation
  • Small language model training for fast retrieval
  • Auto-clustering pipeline
  • Ontology-based knowledge organization

Technical Achievements

1. Portable Deployment

Full infrastructure deployment in 10 minutes with auto-fixing

2. Associative Memory

Fast (<100ms) retrieval using small language models trained on ontology schema

3. Auto-Clustering

HDBSCAN + Cerebras/Llama 3.1 for discovering hidden relationships

4. Auto-Maintenance

Vault health scanner that fixes violations, enriches content, finds relationships

5. Self-Fixing

Deployment system that monitors, detects issues, and fixes them automatically

Why It Matters

Alfred represents several important trends:

1. Self-Deploying AI Systems

Moving from "AI writes code" to "AI deploys and maintains infrastructure"

2. Associative Memory

Solving the context problem with structured vaults and fast retrieval

3. AI-First Business OS

Not just "AI tools" but "AI as operating system" for business

4. Solopreneur Economics

Shifting from SaaS fees to token costs (280x price drop enables this)

5. OpenClaw Extensions

Demonstrating how to build production systems on OpenClaw

From onecontext

Persistent context across sessions - Alfred's vault is a similar concept but more structured

From temporal-io-ai

Workflow orchestration - Alfred uses n8n for automation, Temporal-style reliability

From steve-yegge's Gas Town

Agent orchestration - Alfred is an orchestrated system of agents and automations

From github-spec-kit

Ontology/constitution - Alfred has a Palantir-based ontology for vault structure

Who Built It

David Szabo-Stuban (@ssdavid)

  • Blog: Lumberjack (https://lumberjack.so/)
  • Substack: 7.1K+ subscribers
  • Background: Building AI stuff with no-code, in tech since 2012
  • Previous work: 4 startups, 30+ products shipped
  • Current: Building AI-first business OS for solopreneurs

Open Source

AlfredOS was made open source in August 2025 as a gift for Lumberjack's first birthday.

Deploy: Railway.com template (free tier available)
License: Open source (license not specified in sources)

The "I'm so nerdsniped" Moment

"I'm so nerdsniped."

David's reaction to getting Alfred fully portable captures something important: this is a builder building their dream infrastructure. Not a product, not a startup, just pure "I wanted this to exist so I built it."

That authenticity - building for yourself first - often leads to the most interesting systems.

Future Directions

Based on the journey so far:

  1. LibreChat API - Once shipped, enable true automations combining n8n + agent
  2. Hetzner Migration - Move from Railway to cheaper hosting
  3. API Access - MCP connection to plug Alfred into Claude or other agents
  4. Community Forks - Open source enables others to extend Alfred

Should You Try It?

Yes, if:

  • You're a solopreneur tired of SaaS fees
  • You want to own your business data
  • You're comfortable with self-hosting
  • You want an AI agent that knows all your tools

Wait, if:

  • You need enterprise features
  • You want zero maintenance (this is v1.1, expect bugs)
  • You're not technical enough for self-hosting
  • You prefer SaaS simplicity

Key Quote

"The value of AlfredOS is not the agent. It's the plumbing."

This captures the essence: Alfred is infrastructure for AI-first business, not just another chatbot.

Edit this page