Pack System

What Are Packs?

Packs specialize Working Mind for a domain. A pack is a directory containing a system prompt, tool declarations, skills, personas, and commands. The core agent is pack-agnostic -- it knows nothing about any domain. Packs define the domain.

The Starter Pack

The only pack that ships with Working Mind today. It provides:

• Persistent memory -- built-in SQLite knowledge graph with FTS5 search, contradiction detection, and temporal validity
• Web search -- via Brave Search MCP server (requires BRAVE_API_KEY)
• Web scraping -- via Firecrawl MCP server (requires FIRECRAWL_API_KEY)
• Session summary and export -- built-in curation prompts

The starter pack works immediately. Memory works without any API keys or MCP servers. Search and scraping are optional -- add keys when you're ready.

Pack Anatomy

my-pack/
  pack.json       -- name, version, MCP server declarations, settings
  prompt.md       -- the domain expert system prompt
  skills/         -- reusable subroutines (step-by-step workflows)
  commands/       -- slash command entry points (markdown files)
  personas/       -- execution modes (role-specific behaviors)
  curation/       -- output templates (summarize + export formats)

Every file is human-readable. The directory is forkable (copy it, edit prompt.md), versionable (pack.json has a version field), distributable (push to git), and composable (load multiple packs).

Pack Sections

Each part of the pack serves a specific purpose:

pack.json

{
  "name": "my-pack",
  "version": "0.1.0",
  "description": "A pack for my domain",
  "prompt": "prompt.md",
  "mcpServers": {
    "brave-search": {
      "package": "@brave/brave-search-mcp-server",
      "required": false
    }
  },
  "personas": {
    "analyst": { "prompt": "personas/analyst.md" }
  },
  "skills": {
    "deep-research": { "instructions": "skills/deep-research.md" }
  }
}

Loading Packs

# Load the default starter pack
wmind

# Load a specific pack
wmind --pack researcher

# Load a pack from a local directory
wmind --pack ./my-custom-pack

# Load multiple packs
wmind --pack starter --pack my-custom-pack

Building a Pack

1. Copy the starter pack -- it's in packs/starter/ of the Working Mind repository
2. Edit prompt.md -- write a domain expert system prompt
3. Add MCP servers -- declare them in pack.json under mcpServers
4. Add skills -- create markdown files in skills/
5. Add personas -- create markdown files in personas/
6. Test -- run wmind --pack ./path/to/my-pack

Why Packs Work: The Expert-Curated Advantage

The pack system is built on a specific hypothesis: domain experts who curate their own knowledge graphs produce the highest quality training data that exists. Not synthetic data. Not crowdsourced annotations. Not GPT-4-generated pairs. Expert-curated, iterated-to-correctness, graph-structured knowledge.

The Loop

1. You are a domain expert. You know your field.
2. You instruct Working Mind. It saves what you teach it.
3. You ask questions. It answers from the graph. You know immediately if the answer is right.
4. If it's wrong, you fix the graph. The next answer is right.
5. When the answers are consistently perfect, the graph is done.

That graph -- verified by the only person qualified to verify it -- is the single best source of domain training data. Every entity was explicitly created. Every observation was verified. Every relation was intentional. No other data pipeline produces this quality.

From Graph to Fine-Tuned Model

The pack hypothesis extends further: if the graph produces perfect answers for a domain expert, it contains enough structure to fine-tune a small model that answers those same questions without the graph.

Research supports this:

The key insight: shit in, shit out cannot be prevented by any pipeline. But a domain expert who iterates until the graph is right eliminates the problem at the source. The data quality is solved by definition -- not by validation, not by filtering, not by synthetic quality checks, but by the expert who built it.

When your graph produces perfect answers, you export it. The .oexp format carries the full graph, taxonomy, persona lenses, and readiness assessment. From there, instruction pairs are generated and a small model is fine-tuned. The result: a domain champion that knows your field better than any frontier model -- because it was trained on knowledge you verified yourself.

What We Know and What's Next

What works now:

• The pack format loads correctly
• MCP servers wire up as declared
• Skills activate and inject instructions
• Personas change agent behavior
• Domain-specific packs measurably improve tool selection and task completion
• Expert users who iterate get consistent, correct answers from their graphs

What's in development:

• The graph-to-fine-tune pipeline (.oexp export + instruction pair generation)
• Persona-driven export (same graph, multiple training data flavors)
• Readiness checks (is your graph deep enough for fine-tuning?)

The path:

• Today: build your graph, verify your answers, compound your knowledge
• Next: export your graph as training data
• Future: fine-tune a domain champion from your expertise

Research Context

The pack architecture draws on and aligns with several active research threads: