vibeship-mind 2.3.0

File-based memory system for AI coding assistants

Mind

Mind gives Claude a mind - not just memory across sessions, but focus within them. It remembers what worked, what didn't, and what it's supposed to be building.

When you're vibe coding with Claude, it forgets everything between sessions. What you decided, what broke, what worked - gone. Even worse, in long sessions it starts suggesting the same failed fixes over and over.

Mind fixes both problems with two-layer memory:

• MEMORY.md - Long-term memory across sessions
• SESSION.md - Short-term focus within a session

---

Why Mind?

3 steps. Install, init, connect. Done.

Fully automated. Memory just works - no commands required:

• Claude writes memories as it works
• Session gaps auto-detected (30 min)
• Learnings auto-promoted to long-term memory
• Context auto-injected into CLAUDE.md
• Reminders auto-surface when due or when keywords match

Optional tools are there when you want them, but the core memory flow runs hands-free.

Two-layer memory:

• Cross-session recall (MEMORY.md)
• Within-session focus (SESSION.md)

Zero friction. Claude writes to files, MCP reads them lazily. No database, no cloud, no sync issues.

Human-readable. Plain .md files you can open, edit, or git-track anytime.

Open source. See exactly how it works. No black box.

---

Ready to Give Claude a Mind?

3 steps. Zero friction.

1. Install Mind

pip install vibeship-mind

2. Initialize in your project

Open a terminal in your project folder, then run:

python -m mind init

3. Connect to Claude Code

Copy and paste this to Claude:

Add Mind MCP server to my config. Use command "python" with args ["-m", "mind", "mcp"]

Claude will set it up for you. Restart Claude Code after, and Mind will work automatically.

Manual setup (if you prefer)

Add to your MCP config file:

Mac/Linux: ~/.claude.json Windows: %USERPROFILE%\.claude.json

{
  "mcpServers": {
    "mind": {
      "command": "python",
      "args": ["-m", "mind", "mcp"]
    }
  }
}

Save the file, then restart Claude Code.

---

Alternative: Install from source

git clone https://github.com/vibeforge1111/vibeship-mind.git
cd vibeship-mind
uv sync
uv run mind init

MCP config for source install:

{
  "mcpServers": {
    "mind": {
      "command": "uv",
      "args": ["--directory", "/path/to/vibeship-mind", "run", "mind", "mcp"]
    }
  }
}

---

What's New in 2.2.0

• Semantic Search - mind_search() uses TF-IDF similarity, not just keywords
• Loop Detection - Warns when you're about to repeat a rejected approach
• Smart Promotion - Deduplicates memories, links related entries
• Memory <-> Session Flow - Relevant memories surface when logging blockers

---

MCP Tools (12 total)

Core:

Tool	What it does
mind_recall()	Load session context - CALL FIRST every session
mind_log(msg, type)	Log to session or memory (routes by type)

Type routing for mind_log():

• SESSION.md: experience, blocker, assumption, rejected
• MEMORY.md: decision, learning, problem, progress
• SELF_IMPROVE.md: feedback, preference, blind_spot, skill
• Special: reinforce (boosts pattern confidence)

Reading:

Tool	What it does
mind_session()	Check current session state
mind_search(query)	Semantic search past memories
mind_status()	Check memory health
mind_reminders()	List pending reminders

Actions:

Tool	What it does
mind_blocker(desc)	Log blocker + auto-search memory for solutions
mind_remind(msg, when)	Set reminder - time or context-based
mind_reminder_done(index)	Mark a reminder as done
mind_edges(intent)	Check for gotchas before risky code
mind_checkpoint()	Force process pending memories
mind_add_global_edge()	Add cross-project gotcha

---

How It Works

Long-term (MEMORY.md): Permanent knowledge - decisions, learnings, problems, progress. mind_recall() loads this as context each session.

Short-term (SESSION.md): Working memory buffer:

• Experience - Raw moments, thoughts, what's happening
• Blockers - Things stopping progress
• Rejected - What didn't work and why
• Assumptions - What you're assuming true

When a new session starts (30 min gap), valuable items get promoted from SESSION.md to MEMORY.md automatically.

See docs/HOW_IT_WORKS.md for the full architecture.

---

Reminders

Mind supports two types of reminders:

Time-based:

• "tomorrow", "in 3 days", "next session", "2025-12-20"

Context-based:

• "when I mention auth" - triggers when relevant keywords come up
• "when we work on database" - Claude sees keywords and surfaces reminder naturally

Example: "Remind me to check the security audit when we work on auth"

Reminders are stored in .mind/REMINDERS.md and shown in mind_recall() output.

---

Initialize Mind in Any Project

Open a terminal in your project folder and run:

python -m mind init

That's it! This creates .mind/MEMORY.md and .mind/SESSION.md.

---

Quick Commands

# Check if everything is working
python -m mind doctor

# See what Mind extracted from your notes
python -m mind parse

# Check project status
python -m mind status

# List all registered projects
python -m mind list

---

The Problem This Solves

Across sessions:

• "What did we decide yesterday?" -> Forgotten
• "What gotchas did we hit?" -> Re-discovered every time

Within sessions:

• "Didn't we already try that?" -> Suggests same failed fix 3 times
• "What are we building again?" -> Drifts into rabbit holes

Mind fixes both. Two layers of memory, zero friction.

---

File Structure

your-project/
├── .mind/
│   ├── MEMORY.md     <- Long-term memory (persists)
│   ├── SESSION.md    <- Short-term focus (cleared each session)
│   ├── REMINDERS.md  <- Time and context-based reminders
│   ├── config.json   <- Feature flags for experiments
│   └── state.json    <- Timestamps for session detection
└── CLAUDE.md         <- Mind injects context here

---

Troubleshooting

Problem	Fix
"Command not found"	Use python -m mind instead of just mind
Nothing being captured	Use keywords: decided, problem, learned, gotcha
Claude repeating mistakes	Tell Claude: "Check SESSION.md" or "Add to Rejected Approaches"
Need to check health	Run python -m mind doctor

---

Documentation

• Architecture - Technical deep-dive, design principles
• How It Works - Visual diagrams explaining the system
• MCP Tools - Full tool reference with parameters

---

License

Apache 2.0 - See LICENSE for details.

---

Built by @meta_alchemist

A vibeship.co ecosystem project