skillmind 0.6.0

Active Skill Listener & Trainer — structured memory layer for AI coding assistants

SkillMind

Active Skill Listener & Trainer — the structured memory layer for AI coding assistants.

Replace flat markdown memory files with a vector-database-backed system that listens, learns, and surfaces only relevant context per conversation.

🌐 Visit skill-mind.com — Documentation, setup guides, and live demos.

---

The Problem

Current state	SkillMind
15+ flat markdown memory files	One vector DB with semantic search
Everything loaded into context	Only relevant memories surfaced
Manual save/forget	Auto-listen, auto-classify, auto-expire
No deduplication	Automatic dedup + consolidation
No protection	Sanitizer redacts API keys, PII before storing

Quick Start (3 Commands)

# 1. Install
pip install skillmind[pinecone]

# 2. Initialize
skillmind init --backend pinecone

# 3. Import your existing Claude Code memories
skillmind import ~/.claude/projects/*/memory/

That's it. Your memories are now in a vector DB with semantic search.

How It Works

┌────────────┐    ┌────────────┐    ┌─────────────────┐
│  LISTENER   │───▶│  TRAINER    │───▶│  VECTOR STORE    │
│             │    │             │    │                  │
│ • Git hooks │    │ • Classify  │    │ • Chroma (local) │
│ • File watch│    │ • Sanitize  │    │ • Pinecone       │
│ • Convos    │    │ • Dedup     │    │ • Supabase       │
│ • YouTube   │    │ • Merge     │    │ • Qdrant         │
│ • Screen    │    │ • Expire    │    │ • FAISS          │
└────────────┘    └────────────┘    └─────────────────┘
                                            │
                                            ▼
                                    ┌─────────────────┐
                                    │  CLAUDE CODE     │
                                    │  (MCP Server)    │
                                    │  23 tools        │
                                    └─────────────────┘

6 Vector Store Backends

Backend	Best for	Requires
ChromaDB	Solo dev, local, default	pip install skillmind[chroma]
Pinecone	Multi-device cloud sync	API key
Supabase	SQL + vectors, team sharing	Supabase project
Qdrant	Self-hosted or cloud, great filtering	Qdrant server
FAISS	Offline, air-gapped, fastest	pip install skillmind[faiss]
FalkorDB	GraphRAG (graph + vector, multi-hop)	FalkorDB server (falkordb>=1.6)

All backends implement the same interface — switch anytime with zero data loss.

31 MCP Tools for Claude Code

Once installed, Claude Code gets these tools:

Memory CRUD

Tool	What it does
remember	Store memory (auto-classified, sanitized, deduped)
recall	Semantic search across all memories
forget	Delete a memory
update_memory	Edit existing memory
context	Generate focused context for current file/topic
consolidate	Merge duplicates, expire stale, cleanup
memory_stats	Counts by type, backend info
list_memories	Filter by type/topic
import_markdown_memories	Bulk import from Claude Code markdown files

Video & YouTube Learning

Tool	What it does
learn_youtube	Extract knowledge from YouTube video
learn_youtube_channel	Learn from channel's latest videos
learn_video	Learn from local video/screen recording
record_screen	Record screen to MP4
screenshot	Capture screenshot

Custom Patterns

Tool	What it does
add_pattern	Create auto-detection regex (e.g. client names, SEO terms)
list_patterns	Show all custom patterns
remove_pattern	Delete a pattern

Review Queue

Tool	What it does
set_review_mode	Switch: review (queue first), auto (store directly), off
get_review_mode	Show current mode
review_pending	Show queued memories waiting for approval
approve_memory	Approve single memory → store in vector DB
reject_memory	Reject single memory → discard
approve_all_pending	Approve all at once
reject_all_pending	Clear the queue
edit_pending	Edit content/type/topic before approving

Export & Enrichment (Obsidian / OKF)

Tool	What it does
export_obsidian	Export the whole memory system as an Obsidian wiki vault
sync_obsidian	Incremental sync of new/changed memories into the vault
export_okf	Export memories as a portable OKF bundle (vendor-neutral markdown)
import_okf_bundle	Import an OKF bundle back into the store (round-trip safe)
visualize_okf	Render an OKF bundle as a self-contained local HTML knowledge graph (no server)
enrich_source	Pull knowledge from a source (markdown/web), classify & store it

MCP Setup (Claude Code settings.json)

{
  "mcpServers": {
    "skillmind": {
      "command": "python",
      "args": ["-m", "skillmind.mcp.server"],
      "env": {
        "PINECONE_API_KEY": "your-key",
        "SKILLMIND_BACKEND": "pinecone"
      }
    }
  }
}

Install as a Claude Code Plugin

SkillMind ships as a one-command Claude Code plugin — it registers the MCP server for you, no manual settings.json editing:

# In Claude Code:
/plugin marketplace add AntonioBlago/skillmind
/plugin install skillmind@skillmind

This adds the SkillMind marketplace from the GitHub repo and installs the plugin, which wires up the skillmind serve MCP server automatically. Make sure the CLI is on your PATH first (pip install skillmind[mcp]), then restart Claude Code to load the 31 tools.

Memory Types

Type	What	Auto-expires
user	Role, preferences, expertise	No
feedback	Corrections, confirmed approaches	No
project	Deadlines, client context, status	90 days
reference	External URLs, dashboards, wikis	No
skill	Patterns, workflows, how-tos	No

Obsidian Vault Export (Karpathy Wiki Pattern)

Export your entire memory system as an Obsidian vault with interlinked wiki pages, inspired by Andrej Karpathy's LLM wiki approach.

Complete Setup Guide

Step 1: Install SkillMind

# Install with your preferred backend + MCP server
pip install skillmind[pinecone,mcp,youtube]

# Or install everything
pip install skillmind[all]

Step 2: Configure the MCP Server

Add SkillMind to your Claude Code settings.json:

{
  "mcpServers": {
    "skillmind": {
      "command": "python",
      "args": ["-m", "skillmind.mcp.server"],
      "env": {
        "PINECONE_API_KEY": "your-pinecone-key",
        "SKILLMIND_BACKEND": "pinecone",
        "ANTHROPIC_API_KEY": "your-anthropic-key"
      }
    }
  }
}

Or create a .env file in your project root:

PINECONE_API_KEY=your-pinecone-key
SKILLMIND_BACKEND=pinecone
ANTHROPIC_API_KEY=your-anthropic-key

Step 3: Import Existing Memories

If you already have Claude Code markdown memories:

skillmind import

This scans ~/.claude/projects/*/memory/*.md, auto-classifies each memory, deduplicates, and stores them in your vector DB.

Step 4: Install Obsidian (Free)

1. Download from obsidian.md (Windows/Mac/Linux)
2. Install and open it
3. No account required - works 100% offline

Step 5: Export to Obsidian Vault

Via CLI:

# Full export
skillmind export ~/Documents/MyWiki

# Incremental sync (only new memories)
skillmind sync --vault ~/Documents/MyWiki

Via MCP tools (inside Claude Code):

> export all my memories to an Obsidian vault at ~/Documents/MyWiki

Claude Code will call export_obsidian automatically.

Step 6: Open the Vault in Obsidian

1. Open Obsidian
2. Click "Open folder as vault"
3. Select the vault folder (e.g. ~/Documents/MyWiki)
4. Press Ctrl+G to open the knowledge graph
5. Graph groups are pre-configured with colors per category

Step 7: Keep it in Sync

Every time you add new memories (via remember, learn_youtube, etc.), sync them:

skillmind sync

Or via MCP: sync_obsidian

What Gets Generated

MyWiki/
├── .obsidian/             # Pre-configured: graph groups, bookmarks, plugins
├── CLAUDE.md              # Wiki maintenance instructions for Claude
├── raw/                   # Drop raw source material here (articles, PDFs)
└── wiki/
    ├── index.md           # Master index by category and topic
    ├── log.md             # Operation history
    ├── skills/            # Skill memories (green in graph)
    ├── references/        # Reference memories (blue in graph)
    ├── feedback/          # Feedback memories (teal in graph)
    ├── projects/          # Project memories (orange in graph)
    ├── users/             # User profile memories (purple in graph)
    └── topics/            # Topic MOC pages (light blue in graph)

Each wiki page includes:

• YAML frontmatter with Obsidian-native tags, created, updated dates
• [[Wikilinks]] to related memories across categories
• Topic and category backlinks
• #hashtags for quick filtering
• Confidence scores and source metadata

Graph groups (pre-configured colors):

Category	Color	Obsidian Query
Skills	Green	path:wiki/skills
References	Blue	path:wiki/references
Feedback	Teal	path:wiki/feedback
Projects	Orange	path:wiki/projects
Users	Purple	path:wiki/users
Topics (MOC)	Light Blue	path:wiki/topics
Index pages	Yellow	tag:#MOC

Optional: YouTube Learning with Proxy

If YouTube blocks direct connections, add ScraperAPI proxy support:

# In .env
VPN_PROXY_API_KEY=your-scraperapi-key
SCRAPER_Vendor=scraperapi

Then learn from YouTube videos:

skillmind learn-youtube "https://youtube.com/watch?v=..."
skillmind learn-youtube-channel "UCxxxxx" --limit 5

The extracted knowledge is auto-stored as memories and can be synced to your Obsidian vault.

Portable Second Brain (OKF)

Beyond the Obsidian vault, SkillMind can export your whole memory system as an Open Knowledge Format (OKF) bundle — a vendor-neutral folder of markdown concept docs with relative wiki links, an index.md entry point, a log.md changelog, and # Citations provenance. The bundle is portable across tools and round-trip safe: export, edit elsewhere, import back without losing SkillMind metadata.

skillmind export-okf ~/brain-okf            # Export portable bundle
skillmind import-okf ~/brain-okf --dry-run  # Preview an external bundle
skillmind import-okf ~/brain-okf            # Import (classified + deduplicated)
skillmind viz-okf ~/brain-okf --open        # Render a local HTML knowledge graph

Local Knowledge Graph

Turn any OKF bundle into an interactive, self-contained HTML graph — no server, no build step. viz-okf writes a single okf-graph.html into the bundle that renders concepts as nodes and relative wiki links as edges (Cytoscape.js + marked.js via CDN), with search, type filters, click-through navigation, and backlinks. This is a native re-implementation of the OKF viewer concept — it reads the bundle directly, so it stays in line with the no-enrichment_agent constraint above.

skillmind viz-okf ~/brain-okf                       # Write okf-graph.html into the bundle
skillmind viz-okf ~/brain-okf -o graph.html --open  # Custom name + open in browser

SkillMind adopts only the OKF format and bundle/enrichment concepts — it does not vendor Google's enrichment_agent (which is tied to google-adk + BigQuery), so the bundle stays portable and backend-agnostic.

Enrichment Loop

Pull knowledge from external sources, auto-classify it, and store it as memories — the same pipeline as remember, but from a source adapter (markdown or web):

skillmind enrich -t markdown -s ./notes/              # Learn from local markdown
skillmind enrich -t web -s https://example.com/post   # Learn from a web page

Built-in Sanitizer

API keys, emails, phone numbers, IBANs, and PII are automatically redacted before storing:

Input:  "My Pinecone key is pcsk_EXAMPLExxxxxxxxxxxxxxxx..."
Stored: "My Pinecone key is [REDACTED:PINECONE_KEY]"

Configurable allowlists, custom patterns, and name anonymization.

CLI Reference

skillmind init --backend chroma     # Initialize with backend
skillmind setup                     # Interactive first-run setup
skillmind remember "content"        # Store a memory
skillmind recall "query"            # Semantic search
skillmind list                      # List all memories
skillmind list --type feedback      # Filter by type
skillmind forget <id>               # Delete
skillmind import                    # Import Claude Code memories
skillmind migrate --from chroma --to pinecone   # Move data between backends
skillmind consolidate               # Cleanup & deduplicate
skillmind context                   # Generate context for current dir
skillmind stats                     # Show statistics
skillmind serve                     # Start MCP server

# Export / portable second brain
skillmind export ~/Documents/MyWiki         # Export to Obsidian vault
skillmind sync --vault ~/Documents/MyWiki   # Incremental vault sync
skillmind export-okf ~/brain-okf            # Export portable OKF bundle
skillmind import-okf ~/brain-okf            # Import an OKF bundle (--dry-run to preview)
skillmind viz-okf ~/brain-okf --open        # Render a self-contained local HTML graph

# Enrichment loop (source → classify → store)
skillmind enrich -t markdown -s ./notes/    # Learn from markdown files
skillmind enrich -t web -s https://example.com/post   # Learn from a web page

Installation

# Core + one backend
pip install skillmind[chroma]       # Local (default)
pip install skillmind[pinecone]     # Cloud sync
pip install skillmind[supabase]     # SQL + vectors
pip install skillmind[qdrant]       # Self-hosted
pip install skillmind[faiss]        # Offline

# Add-ons
pip install skillmind[youtube]      # YouTube learning
pip install skillmind[video]        # Screen recording + OCR
pip install skillmind[mcp]          # MCP server for Claude Code

# Everything
pip install skillmind[all]

Architecture

src/skillmind/
├── models.py          # Memory, QueryFilter, QueryResult (Pydantic)
├── config.py          # SkillMindConfig (YAML-based)
├── embeddings.py      # EmbeddingEngine (sentence-transformers | OpenAI)
├── trainer.py         # Auto-classify, dedup, merge, consolidate
├── sanitizer.py       # Redact API keys, PII before storage
├── review.py          # Review queue (queue → approve/reject → store)
├── setup.py           # Interactive first-run setup
├── listener.py        # GitListener, FileListener, ConversationListener
├── context.py         # ContextGenerator — dynamic context injection
├── migration.py       # Import existing Claude Code markdown memories
├── enrichment.py      # Enrichment loop: source → classify → store
├── store/
│   ├── base.py            # Abstract MemoryStore interface
│   ├── chroma_store.py    # ChromaDB backend
│   ├── pinecone_store.py  # Pinecone backend
│   ├── supabase_store.py  # Supabase/pgvector backend
│   ├── qdrant_store.py    # Qdrant backend
│   ├── faiss_store.py     # FAISS + JSON backend
│   └── falkordb_store.py  # FalkorDB backend (GraphRAG: graph + vector)
├── sources/
│   ├── base.py            # Source adapter interface
│   ├── markdown_source.py # Markdown files → memories
│   └── web_source.py      # Web pages → memories
├── exporters/
│   ├── obsidian.py        # Obsidian vault exporter (Karpathy wiki pattern)
│   ├── okf.py             # OKF bundle exporter (portable, vendor-neutral)
│   └── okf_viz.py         # OKF → self-contained local HTML knowledge graph
├── importers/
│   └── okf.py             # OKF bundle importer (round-trip safe)
├── video/
│   ├── youtube_learner.py     # YouTube transcript extraction
│   ├── video_learner.py       # Local video learning
│   └── screen_recorder.py     # Screen capture
├── mcp/
│   └── server.py      # FastMCP server (31 tools)
└── cli/
    └── main.py        # CLI (click-based)

Complementary MCP Tools

SkillMind works great alongside other MCP servers:

MCP Server	What it adds	Install
Visibly AI MCP	SEO skills: keyword research, backlinks, site audit, GSC queries, on-page analysis	Hosted MCP — no install; add HTTP config to settings.json
Notion MCP	Read/write Notion pages, databases, todos	Built-in
Playwright MCP	Browser automation, screenshots, web scraping	Built-in

Example: SEO Memory + Skills

# SkillMind installs locally; Visibly AI is a hosted MCP (no install needed)
pip install skillmind[pinecone,mcp]

{
  "mcpServers": {
    "skillmind": {
      "command": "python",
      "args": ["-m", "skillmind.mcp.server"],
      "env": { "SKILLMIND_BACKEND": "pinecone", "PINECONE_API_KEY": "..." }
    },
    "visiblyai": {
      "type": "http",
      "url": "https://mcp.visibly-ai.com/mcp",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }
  }
}

SkillMind remembers your SEO preferences, client keywords, and audit findings. Visibly AI provides the live SEO data. Together: persistent SEO intelligence.

Used by

peec-ai-skills

A Claude Code skill repository for Peec AI (brand-visibility tracking in LLM search) that uses SkillMind as its cross-project memory layer. The skillmind-learner skill writes causal, evidence-backed patterns to SkillMind after a Peec AI growth-loop outcome is measured, and recalls matching patterns as priors for the growth-agent orchestrator on the next run — so lessons learned on project A inform decisions on project B.

See skills/skillmind-learner/SKILL.md for the full integration contract (read vs. write mode, required tags, pattern schema, handoff points from the other 5 Peec skills).

Contributing

PRs welcome! See CONTRIBUTING.md.

License

MIT — see LICENSE.

---

Built by Antonio Blago | skill-mind.com