mindforge-kb 0.3.0

A local semantic memory engine that transforms AI conversation transcripts into a structured, queryable knowledge base.

MindForge

Transform messy AI conversations into a structured, queryable knowledge base.

MindForge is a local-first semantic memory engine. Feed it raw conversation transcripts and it distills them into atomic, interlinked concepts -- complete with a knowledge graph, wiki-style links, and optional vector search. Think of it as a second brain that actually organizes itself.

---

Why MindForge?

Every conversation with an AI produces knowledge. Most of it vanishes into scroll history.

MindForge captures that knowledge and turns it into something you can navigate, query, and build on:

• Concepts, not conversations -- Each idea becomes its own clean Markdown file
• Relationships are explicit -- Wiki-style [[links]] and typed edges (uses, depends_on, enables)
• Knowledge graph included -- Visualize how concepts connect
• Semantic search ready -- Find concepts by meaning, not just keywords
• LLM-powered extraction -- Optionally use Ollama or any OpenAI-compatible API for dramatically better results
• 100% local -- No cloud required. Your knowledge stays yours.

---

Install

Pick the path that matches you:

Python users (recommended)

uv tool install mindforge-kb
# or
pipx install mindforge-kb

With optional embeddings for semantic search:

uv tool install 'mindforge-kb[embeddings]'

Homebrew (macOS / Linux)

brew tap acceleratedindustries/mindforge
brew install mindforge

Single binary (no Python required)

Download the latest binary for your platform from GitHub Releases:

curl -L https://github.com/AcceleratedIndustries/MindForge/releases/latest/download/mindforge-macos-arm64 -o mindforge
chmod +x mindforge
./mindforge --help

From source

git clone https://github.com/AcceleratedIndustries/MindForge.git
cd MindForge
pip install -e ".[dev]"

Quick Start

# Run on your transcripts
mindforge ingest --input path/to/transcripts --output output

# Ask questions
mindforge query "How does semantic search work?"

# See what you've built
mindforge stats

With LLM extraction (recommended for best results):

# Using local Ollama
mindforge ingest --input transcripts/ --llm --llm-model llama3.2

# Using OpenAI
mindforge ingest --input transcripts/ --llm --llm-provider openai --llm-api-key sk-...

# Using any OpenAI-compatible endpoint (vLLM, LM Studio, Together, etc.)
mindforge ingest --input transcripts/ --llm --llm-provider openai \
    --llm-base-url http://localhost:8000 --llm-model my-model

---

What It Produces

Given a directory of conversation transcripts, MindForge outputs:

Concept Files (Markdown)

Each concept gets its own file with YAML frontmatter, a clean definition, explanation, key insights, and wiki-style links to related concepts:

---
title: "KV Cache"
slug: "kv-cache"
tags: [transformers, inference, optimization]
confidence: 0.90
---

# KV Cache

## Definition

KV Cache is a mechanism that stores the Key and Value matrices from the
attention computation of previously processed tokens, avoiding redundant
recomputation during autoregressive generation.

## Key Insights

- Trades memory for computation -- critical trade-off in production LLM serving
- Techniques like Multi-Query Attention reduce KV cache size

## Related Concepts

- [[Vector Embeddings]]
- [[Attention Mechanism]]

Knowledge Graph (JSON)

A NetworkX-powered directed graph with typed edges, exportable as JSON:

{
  "nodes": [{"id": "kv-cache", "label": "KV Cache"}],
  "edges": [{"source": "rag", "target": "semantic-search", "type": "uses"}]
}

Embeddings Index (Optional)

FAISS-backed vector index for semantic search over your concepts.

---

The Pipeline

Transcripts --> Parse --> Chunk --> Extract --> Deduplicate --> Distill --> Link --> Graph
                                     |                                      |
                                     +-- Heuristic (patterns, headings)     +-- Wiki-links
                                     +-- LLM (Ollama / OpenAI) [optional]   +-- Typed relationships

Stage	What it does
Parse	Multi-format transcript parser (role-prefixed, heading-style, separators)
Chunk	Semantic chunking that respects paragraphs, headings, and code blocks
Extract	Identify concepts via definition patterns, heading analysis, keyword frequency -- or LLM
Deduplicate	Merge near-duplicates by slug matching and Jaccard similarity
Distill	Clean conversational fluff, extract insights, build structured definitions
Link	Detect relationships via co-occurrence, keyword overlap, and structural patterns
Graph	Build a NetworkX knowledge graph, export as JSON
Index	Optional FAISS + sentence-transformers for semantic queries

---

Project Structure

mindforge/
├── cli.py                  # CLI: ingest / query / stats
├── config.py               # Central configuration
├── pipeline.py             # 6-stage pipeline orchestrator
├── ingestion/
│   ├── parser.py           # Multi-format transcript parser
│   ├── chunker.py          # Semantic chunking
│   └── extractor.py        # Heuristic concept extraction
├── llm/
│   ├── client.py           # Ollama + OpenAI HTTP client (stdlib only)
│   ├── extractor.py        # LLM-based concept extraction
│   └── distiller.py        # LLM-aware concept distillation
├── mcp/
│   └── server.py           # MCP server (JSON-RPC over stdio)
├── distillation/
│   ├── concept.py          # Data models (Concept, Relationship, ConceptStore)
│   ├── deduplicator.py     # Similarity-based deduplication
│   ├── distiller.py        # Raw -> clean concept transformation
│   └── renderer.py         # Concept -> Markdown with frontmatter
├── linking/
│   └── linker.py           # Relationship detection + wiki-links
├── graph/
│   └── builder.py          # NetworkX graph + JSON export
├── embeddings/
│   └── index.py            # FAISS + sentence-transformers index
├── query/
│   └── engine.py           # Keyword + semantic search
└── utils/
    └── text.py             # Slugify, similarity, keyword extraction

---

Installation

Core (no external dependencies beyond NetworkX):

pip install -e .

With semantic search:

pip install -e ".[embeddings]"

For development:

pip install -e ".[dev]"
pytest

---

Transcript Formats

MindForge accepts Markdown or plain text files. Drop them in a directory and point --input at it.

Supported conversation formats:

• Role-prefixed: User: ... / Assistant: ...
• Heading-style: ## User / ## Assistant
• Separator-based: Turns separated by ---
• Plain text: Treated as a single knowledge document

---

Heuristic vs LLM Extraction

MindForge works in two modes:

	Heuristic (default)	LLM-assisted (--llm)
Speed	Instant	Depends on model
Dependencies	None	Ollama or API
Quality	Good for well-structured transcripts	Excellent for any text
Relationships	Detected via patterns	Explicitly identified by the LLM
Fallback	N/A	Automatic fallback to heuristic

When --llm is enabled, MindForge runs both extractors and merges the results. LLM-identified concepts take priority, and any unique heuristic findings are added in. If the LLM server is unreachable, it falls back to heuristic-only with a clear message.

---

Relationship Types

MindForge tracks eight types of concept relationships:

Type	Meaning	Example
uses	A uses B	RAG uses Semantic Search
depends_on	A requires B	Semantic Search depends on Embeddings
enables	A makes B possible	Embeddings enables Similarity Search
improves	A enhances B	Hybrid Search improves Recall
part_of	A is a component of B	HNSW part of Vector Database
example_of	A is an instance of B	Qdrant example of Vector Database
contrasts_with	A differs from B	Keyword Search contrasts with Semantic Search
related_to	General association	KV Cache related to Attention Mechanism

---

MCP Server (AI Agent Interface)

MindForge includes an MCP (Model Context Protocol) server that lets external AI agents query your knowledge base as a tool.

# Start the MCP server
mindforge mcp --output path/to/knowledge-base

Available tools:

Tool	Description
search	Natural language search across all concepts
get_concept	Get full details of a concept by name or slug
list_concepts	List all concepts, optionally filtered by tag
get_neighbors	Get related concepts via the knowledge graph
get_stats	Knowledge base statistics and most central concepts

Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "mindforge": {
      "command": "mindforge",
      "args": ["mcp", "--output", "/path/to/your/output"]
    }
  }
}

---

Roadmap

• MCP server interface -- Expose the knowledge base as a tool server for external AI agents
• Incremental ingestion -- Content hashing to skip already-processed transcripts
• Concept versioning -- Track how concepts evolve across ingestion runs
• Confidence decay -- Unreinforced concepts fade over time, surfacing review candidates
• Auto-refactoring -- Merge or split concepts as the knowledge base grows
• Graph visualization -- Interactive web UI for exploring the knowledge graph

---

License

Business Source License 1.1 (BUSL-1.1). See LICENSE for details.

In short: free for any use, including commercial use, except offering MindForge as a hosted or managed service with functionality substantially similar to a paid offering by Accelerated Industries. Each release auto-converts to the Apache License 2.0 on April 21, 2028.