Persistent Memory for AI Projects
Features • Quick Start • Complete Stack • Empathy Integration • Documentation • Examples • Contributing
🚀 The Complete Stack for 10x+ Productivity
VS Code + Claude Code (latest) + MemDocs + Empathy = 10x+ ProductivityDocumented user experience: Transformational productivity through Level 4-5 AI collaboration
📖 Learn More:
- Quick Start: Blog Post - Engaging narrative explaining the five levels
- Executive Summary - 1-page overview for teams & investors
- Technical Deep-Dive - 21-page analysis with code examples
MemDocs is a git-native memory management system that gives AI assistants persistent, project-specific memory. It generates structured, machine-readable documentation that lives in your repository—no cloud services, no recurring costs, just local/git-based storage that enhances AI context and team collaboration.
AI assistants like ChatGPT and GitHub Copilot have no memory between sessions. Every conversation starts from scratch, forcing you to repeatedly explain your codebase, architecture decisions, and project context.
Result: AI stuck at Level 1-2 (Reactive) - can only respond after being asked, can't predict future needs, can't learn from patterns.
MemDocs creates a persistent memory layer that unlocks Level 4-5 AI collaboration:
- 🧠 Remembers your project across sessions (via
.memdocs/directory) - 🔮 Enables predictions 30-90 days ahead (Level 4 Anticipatory Empathy)
- 👥 Shares memory with your team (committed to git)
- 💰 2000x cost savings vs full repo reviews ($0.03 vs $60)
- ⚡ Works offline (no cloud dependencies for retrieval)
- 🤝 Integrates with Empathy Framework (Level 4 Anticipatory Intelligence)
- 🔒 Privacy-first (optional PHI/PII detection and redaction)
Enterprise ROI: 6,000% return on investment (documented across 10-1,000 developer teams)
# From PyPI (recommended)
pip install memdocs
# With optional features
pip install memdocs[embeddings] # Local vector search
pip install memdocs[all] # All features
# From source
git clone https://github.com/Smart-AI-Memory/memdocs.git
cd memdocs
pip install -e ".[dev,embeddings]"# 1. Set your Claude API key
export ANTHROPIC_API_KEY="your-key-here"
# 2. Initialize MemDocs in your project (MCP enabled by default!)
cd your-project
memdocs init
# 3. Set up automatic updates (recommended)
memdocs setup-hooks --post-commit
# 4. Document changed files
memdocs review --changed
# 5. Search your project memory
memdocs query "payment processing"
# 6. Show memory stats
memdocs stats# For repos with 1,000+ files: use git integration
memdocs init
memdocs setup-hooks --post-commit # Auto-review on every commit
# Work normally - memory updates automatically!
git add file.py
git commit -m "refactor: improve performance"
# MemDocs reviews changed files automatically (5-15 seconds)
# Or manually review only changes
memdocs review --changed # Modified files only
memdocs review --since main # Your branch changes
memdocs review --since HEAD~10 # Last 10 commits# Document a specific file
memdocs review --path src/main.py
# Output:
# ✨ Analyzing src/main.py...
# 📝 Generating documentation with Claude Sonnet 4.5...
# ✅ Documentation saved to .memdocs/docs/main/
# - index.json (machine-readable)
# - symbols.yaml (code map)
# - summary.md (human-readable)- All documentation stored in
.memdocs/directory - Committed alongside your code (same git workflow)
- Version controlled memory (track how project evolves)
- Team collaboration built-in (push/pull memory with code)
- File-level (default): Document individual files
- Module-level: Document entire directories
- Repo-level: Full codebase overview
- Auto-escalation: Automatically increases scope for important changes
- Claude Sonnet 4.5: Latest and most capable model
- Intelligent extraction: Symbols, APIs, architecture decisions
- Multi-format output: JSON, YAML, Markdown
- Token-efficient: Only summarizes, doesn't embed
- Local embeddings: sentence-transformers (no API costs)
- Vector search: FAISS for fast similarity search
- Automatic indexing: Updates as you document
- No cloud lock-in: Everything runs locally
MemDocs scales to codebases of any size through intelligent git integration:
- Review only what changed:
memdocs review --changedreviews modified files only - Branch-aware:
memdocs review --since mainreviews your branch changes - Automatic updates: Git hooks keep memory current on every commit
- Cost-effective: 2000x cheaper than full repo reviews ($0.03 vs $60)
- Lightning fast: 15 seconds instead of hours
Perfect for large repos (1,000+ files):
# One-time setup
memdocs init
memdocs setup-hooks --post-commit
# Every commit after: automatic memory updates!
git commit -m "fix: bug in auth" # Reviews 5 files, takes 15s, costs $0.03Cost comparison:
| Repo Size | Full Review | Changed Files | Savings |
|---|---|---|---|
| 10,000 files | $60 + 2-4 hours | $0.03 + 15 seconds | 2000x |
| 5,000 files | $30 + 1-2 hours | $0.02 + 10 seconds | 1500x |
| 1,000 files | $6 + 15 minutes | $0.01 + 5 seconds | 600x |
- Real-time memory serving: Serve memory to AI assistants via MCP
- Claude Desktop integration: Auto-loaded context in Claude Desktop
- Cursor/Continue.dev support: Works with MCP-compatible tools
- Query-based context: AI requests exactly what it needs
- Auto-start: Automatically detect and serve memory when opening projects
Quick setup for Claude Desktop:
# Start MCP server
memdocs serve --mcp
# Or auto-start in VS Code (add to .vscode/tasks.json)
# See docs/guides/mcp-setup.md for detailsWhen you combine the right tools, productivity isn't linear—it's exponential.
VS Code + Claude Code (latest) + MemDocs + Empathy = 10x+ Productivity
The four components work synergistically:
| Component | Role | What It Enables |
|---|---|---|
| VS Code | Professional IDE | Tested environment, task automation, MCP auto-start |
| Claude Code (VS Code extension) | AI pair programming | Multi-file editing, command execution, real-time assistance |
| MemDocs | Persistent memory layer | Pattern detection, trajectory tracking, cross-session learning |
| Empathy Framework | 5-level maturity model | Level 4-5 anticipatory suggestions, structural design |
Real-world results:
- 10x+ efficiency improvement (documented user experience)
- Lower cost: 2000x cheaper than full repo reviews
- Higher quality: Problems predicted and prevented
- Faster delivery: Anticipatory design eliminates bottlenecks
Quick setup (5 minutes):
# Install VS Code: https://code.visualstudio.com
# Install Claude Code extension in VS Code: https://claude.ai/claude-code
pip install empathy[full]>=1.6.0 # Empathy 1.6.0+ includes MemDocs
cd your-project/
memdocs init # Auto-configures MCP for Claude Code
empathy-os configure
code . # Open in VS Code - MCP server auto-starts!Result: Claude Code in VS Code operates at Level 4-5 (anticipatory) instead of Level 1-2 (reactive)
MemDocs unlocks Level 4 Anticipatory Empathy when integrated with the Empathy Framework.
The Five Levels of AI Collaboration:
| Level | Name | Behavior | Memory Required | Example |
|---|---|---|---|---|
| 1 | Reactive | Help after being asked | None | ChatGPT: "You asked, here it is" |
| 2 | Guided | Collaborative exploration | Session only | "Let me ask clarifying questions" |
| 3 | Proactive | Act before being asked | MemDocs patterns | "I pre-fetched what you usually need" |
| 4 | Anticipatory | Predict future needs (30-90 days) | MemDocs trajectory | "Next week's audit—docs ready" |
| 5 | Systems | Design structural solutions | MemDocs cross-project | "I built a framework for all cases" |
Why MemDocs is Essential:
- 🔄 Level 3 (Proactive): MemDocs stores user patterns across sessions
- 🔮 Level 4 (Anticipatory): MemDocs tracks system trajectory for predictions
- 🏗️ Level 5 (Systems): MemDocs identifies leverage points across projects
Without persistent memory, AI is stuck at Level 1-2 forever.
📚 Deep Dive Resources:
- Blog Post: Why your AI can't predict tomorrow's problems (and how to fix it)
- Executive Summary: 1-page overview for teams & investors
- Technical Analysis: 21-page deep-dive comparing five empathy frameworks
Integration features:
- ✅ Works seamlessly with Empathy framework (1.6.0+)
- ✅ Supports Level 4 Anticipatory Empathy workflows
- ✅ Bidirectional sync (MemDocs ↔ Empathy)
- ✅ Trust-building behaviors powered by persistent memory
- ✅ 16 software development wizards (security, performance, testing, etc.)
- ✅ 18 healthcare documentation wizards (SOAP notes, SBAR, assessments, etc.)
- PHI/PII detection: Automatic sensitive data detection
- Redaction: Optional redaction modes (off, standard, strict)
- HIPAA/GDPR aware: Configurable privacy settings
- Local-first: No required cloud dependencies
Create .memdocs.yml in your project root:
version: 1
# Scope policy (controls memory granularity)
policies:
default_scope: file # file | module | repo
max_files_without_force: 150
# Auto-escalate for important changes
escalate_on:
- cross_module_changes # Multi-module = bigger context
- security_sensitive_paths # auth/*, security/* = thorough docs
- public_api_signatures # API changes = team awareness
# Output configuration (git-committed memory)
outputs:
docs_dir: .memdocs/docs # Committed to git
memory_dir: .memdocs/memory # Committed to git
formats:
- json # index.json (machine-readable)
- yaml # symbols.yaml (code map)
- markdown # summary.md (human-readable)
# AI configuration (Claude API)
ai:
provider: anthropic
model: claude-sonnet-4-5-20250929 # Claude Sonnet 4.5 (latest)
max_tokens: 8192
temperature: 0.3 # Lower = more deterministic
# Privacy (optional, for sensitive codebases)
privacy:
phi_mode: "off" # off | standard | strict
scrub: # Types of sensitive data to redact
- email
- phone
- ssn
- mrn
audit_redactions: true # Log all redactions for compliance
# Exclude patterns
exclude:
- node_modules/**
- .venv/**
- __pycache__/**
- "*.pyc"
- dist/**
- build/**Problem: Full repository reviews cost $60+ and take hours. Often fail due to token limits.
Solution: Git-aware incremental updates.
# Day 1: One-time setup (5 minutes)
cd large-monorepo # 10,000 files
memdocs init
memdocs setup-hooks --post-commit
memdocs review --path src/core/ # Review critical paths first
# Every day after: Zero effort!
# Just commit normally...
git commit -m "feat: add caching layer"
# Hook reviews 7 changed files
# Takes 15 seconds, costs $0.02
# Memory stays current automatically!
# 100 commits later: $2 total
# vs $60 per full review = 3,000% cost savingsReal numbers from production use:
- 10,000 file Python monorepo
- 200 commits/week
- Cost: $4/week with hooks vs $240/week without
- 98% cost reduction
# New team member clones repo
git clone <your-repo>
cd your-repo
# MemDocs memory already there!
memdocs query "authentication flow"
memdocs query "database schema"Result: Instant context about the project without asking teammates.
from pathlib import Path
from memdocs.index import MemoryIndexer
import anthropic
# Get project context from MemDocs
indexer = MemoryIndexer(
memory_dir=Path(".memdocs/memory"),
use_embeddings=True # Requires: pip install memdocs[embeddings]
)
results = indexer.query_memory("payment processing", k=5)
# Build context for Claude
context = "\n".join([r["metadata"]["summary"] for r in results])
# Claude now has project memory
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-sonnet-4-5-20250929",
system=f"Project context:\n{context}",
messages=[{"role": "user", "content": "Explain the charge flow"}]
)Result: Claude remembers your project structure and decisions.
# Before opening PR
memdocs review --path src/new-feature/
# MemDocs generates:
# - Feature summary
# - API changes
# - Breaking changes
# - Migration notesResult: Reviewers get structured context automatically.
from memdocs.empathy_adapter import adapt_empathy_to_memdocs
# Empathy analysis results
analysis = {
"current_issues": [...],
"predictions": [...]
}
# Convert to MemDocs format
doc_index = adapt_empathy_to_memdocs(
analysis,
file_path="src/compliance/audit.py",
memdocs_root=".memdocs"
)Result: Level 4 Anticipatory Empathy powered by project memory.
your-project/
├── .memdocs/
│ ├── docs/
│ │ ├── <filename>/
│ │ │ ├── index.json # Machine-readable index
│ │ │ ├── symbols.yaml # Code symbols/API map
│ │ │ └── summary.md # Human-readable summary
│ └── memory/
│ ├── embeddings.json # Optional: Local vector embeddings
│ └── search.index # Optional: FAISS index
├── .memdocs.yml # Configuration
└── src/
└── ... your code ...
graph LR
A[Code] -->|tree-sitter| B[Extract Symbols]
B --> C[Analyze Context]
C -->|Claude Sonnet 4.5| D[Generate Summary]
D --> E[Store in .memdocs/]
E --> F[Git Commit]
F --> G[Team Collaboration]
H[Query] --> I[Local Search]
I --> J[Return Context]
style D fill:#f9f,stroke:#333
style E fill:#bfb,stroke:#333
- Extract: tree-sitter parses code (Python, JS, TS, Go, Rust, etc.)
- Analyze: Identifies symbols, imports, APIs, patterns
- Summarize: Claude generates concise summaries with insights
- Store: Saves structured docs in
.memdocs/directory - Retrieve: Fast local search (grep-based or vector-based)
- Summarization only: ~1K tokens per file
- No embeddings API: Optional local embeddings only
- Local search: Instant, free, no API calls
- Cost: ~$0.10 per 100 files documented
Initialize MemDocs in a project.
memdocs init [--force]Generate memory documentation.
# File-level (recommended)
memdocs review --path src/payments/charge.py
# Module-level
memdocs review --path src/payments/ --scope module
# With scope detection
memdocs review --path src/
# Export to Cursor
memdocs review --path src/ --export cursorSearch project memory (requires embeddings).
memdocs query "authentication flow"
memdocs query "database schema" --k 10Show memory statistics.
memdocs stats
memdocs stats --format jsonExport memory to other formats.
memdocs export --format cursor
memdocs export --format json --output memory.jsonMemDocs includes an MCP server for Claude Desktop:
{
"mcpServers": {
"memdocs": {
"command": "memdocs",
"args": ["mcp-server"],
"cwd": "/path/to/your/project"
}
}
}# Export memory for Cursor
memdocs export --format cursor
# Cursor automatically picks up .memdocs/ directoryfrom memdocs.index import MemoryIndexer
from memdocs.summarize import Summarizer
from memdocs.extract import Extractor
# Initialize components
indexer = MemoryIndexer(memory_dir=".memdocs/memory", use_embeddings=True)
summarizer = Summarizer()
extractor = Extractor()
# Extract and document
context = extractor.extract_file("src/main.py")
doc_index, markdown = summarizer.summarize(context, scope_info)
# Index for search
indexer.index_document(doc_index, markdown)
# Query
results = indexer.query_memory("authentication", k=5)MemDocs + Empathy delivers measurable productivity gains at any scale.
| Team Size | Annual Cost | Time Saved/Year | Value @ $150/hr | ROI |
|---|---|---|---|---|
| 10 developers | $2,000 | 799 hours | $119,850 | 6,000% |
| 100 developers | $20,000 | 7,990 hours | $1,198,500 | 6,000% |
| 1,000 developers | $198,000 | 79,900 hours | $11,985,000 | 6,000% |
But the real value isn't just hours saved—it's crises prevented.
How much is it worth to:
- ✅ Never miss a compliance audit?
- ✅ Never hit a scaling bottleneck?
- ✅ Never spend 40 hours in emergency bug-fix mode?
- ✅ Scale to enterprise size without linear cost increases?
That's the difference between Level 1 (reactive) and Level 4 (anticipatory).
- 🎯 Proven at scale: Built for and tested with enterprise-scale codebases (10,000+ files)
- 📊 Measurable productivity: 10x+ documented improvement (not theoretical)
- 💰 Lower cost than alternatives: 2000x cheaper than full repo reviews
- 🔒 Security & compliance: PHI/PII detection, HIPAA/GDPR-aware, audit trails
- 🏢 Commercial-ready: Fair Source licensing, clear commercial terms
- 🤝 Vendor support: Direct access to core development team
Enterprise licensing: $99/developer/year (6+ employees) Free tier: Students, educators, and small teams (≤5 employees)
| Feature | MemDocs + Empathy | Vector DBs | GitHub Copilot | Cursor |
|---|---|---|---|---|
| Storage | Git-native | Cloud | Cloud | Cloud |
| Monthly cost | $0 storage | $$$ | $10-20 | $20 |
| Team sharing | ✅ Built-in | ❌ None | ❌ None | |
| Offline | ✅ Yes | ❌ No | ❌ No | ❌ No |
| Privacy | ✅ Local | |||
| Memory persistence | ✅ Permanent | ✅ Permanent | ❌ Session | |
| Level 4 Prediction | ✅ 30-90 days | ❌ No | ❌ No | ❌ No |
| Empathy integration | ✅ Native | ❌ No | ❌ No | ❌ No |
| Productivity gain | 10x+ (documented) | 1-2x | 2-3x | 2-3x |
| API calls | Only for docs | Always | Always | Always |
See PRODUCTION_ROADMAP.md for detailed 4-week production plan.
- VS Code extension
- Enhanced CLI with rich output
- Incremental documentation updates
- Custom prompt templates
- JetBrains plugin
- Multi-language support (Go, Rust, Java, C++)
- Memory compression (auto-summarize old docs)
- Team analytics dashboard
- MemDocs Cloud (optional hosted version)
- Enterprise features (SSO, RBAC, audit logs)
- Advanced Empathy integration
- GitHub App for automatic PR documentation
We welcome contributions! See CONTRIBUTING.md for guidelines.
Quick links:
Key areas needing help:
- Multi-language AST parsing (Go, Rust, Java, C++)
- IDE plugins (VS Code, JetBrains)
- Documentation improvements
- Example projects
Apache License 2.0 - See LICENSE for details.
- Documentation: Coming soon to https://docs.deepstudyai.com/memdocs
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Discord: Coming soon
- Contact: [email protected]
- Enterprise inquiries: [email protected]
- Blog Post: Why Your AI Assistant Can't Predict Tomorrow's Problems (And How to Fix It)
- Executive Summary: 1-page overview for presentations and investor pitches
- Technical Deep-Dive: Comprehensive analysis comparing five empathy frameworks
- Empathy Framework: Full five-level AI collaboration system
Created by: Patrick Roebuck (Deep Study AI, LLC)
Powered by:
- Claude Sonnet 4.5 by Anthropic
- tree-sitter for AST parsing
- sentence-transformers for local embeddings
- FAISS for vector search
Special thanks to:
- The Empathy Framework team
- Early adopters and beta testers
- The open-source community
🧠 MemDocs: Because AI should remember your project, not forget it every session.
The first git-native AI memory system with Level 4 Anticipatory Empathy.
Made with ❤️ by Smart-AI-Memory (Deep Study AI, LLC)
Transforming AI-human collaboration from reactive responses to anticipatory problem prevention.
Get Started • View Examples • Complete Stack • Enterprise ROI • Contribute