SuperLocalMemory V2

Your AI Finally Remembers You

⚡ Created & Architected by Varun Pratap Bhardwaj ⚡
Solution Architect • Original Creator • 2026

Stop re-explaining your codebase every session. 100% local. Zero setup. Completely free.

superlocalmemory.com • Quick Start • Why This? • Features • Docs • Issues

Created by Varun Pratap Bhardwaj • 💖 Sponsor • 📜 Attribution Required

---

What's New in v2.7 — "Your AI Learns You"

SuperLocalMemory now learns your patterns, adapts to your workflow, and personalizes recall — all 100% locally on your machine. No cloud. No LLM. Your behavioral data never leaves your device.

Adaptive Learning System

Your memory system evolves with you through three learning layers:

Layer	What It Learns	How
Tech Preferences	"You prefer FastAPI over Django" (83% confidence)	Cross-project frequency analysis with Bayesian scoring
Project Context	Detects your active project from 4 signals	Path analysis, tags, profile, content clustering
Workflow Patterns	"You typically: docs → architecture → code → test"	Time-weighted sliding-window sequence mining

Three-Phase Adaptive Ranking

Recall results get smarter over time — automatically:

1. Phase 1 (Baseline): Standard search — same as v2.6
2. Phase 2 (Rule-Based): After ~20 feedback signals — boosts results matching your preferences
3. Phase 3 (ML Ranking): After ~200 signals — LightGBM LambdaRank re-ranks with 9 personalized features

Privacy by Design — GDPR Compliant

Concern	SuperLocalMemory v2.7	Cloud-Based Alternatives
Where is learning data?	~/.claude-memory/learning.db on YOUR machine	Their servers, their terms
Who processes your behavior?	Local gradient boosting (no LLM, no GPU)	Cloud LLMs process your data
Right to erasure (GDPR Art. 17)?	slm learning reset — one command, instant	Submit a request, wait weeks
Data portability?	Copy the SQLite file	Vendor lock-in
Telemetry?	Zero. Absolutely none.	Usage analytics, behavior tracking

Your learning data is stored separately from your memories. Delete learning.db and your memories are untouched. Delete memory.db and your learning patterns are untouched. Full data sovereignty.

Research-Backed Architecture

Every component is grounded in peer-reviewed research, adapted for local-first operation:

Component	Research Basis
Two-stage retrieval pipeline	BM25 → re-ranker (eKNOW 2025)
Adaptive cold-start ranking	Hierarchical meta-learning (LREC 2024)
Time-weighted sequence mining	TSW-PrefixSpan (IEEE 2020)
Bayesian confidence scoring	MACLA (arXiv:2512.18950)
LightGBM LambdaRank	Pairwise ranking (Burges 2010, MO-LightGBM SIGIR 2025)
Privacy-preserving feedback	Zero-communication design — stronger than differential privacy (ADPMF, IPM 2024)

New MCP Tools

Tool	Purpose
memory_used	Tell the AI which recalled memories were useful — trains the ranking model
get_learned_patterns	See what the system has learned about your preferences
correct_pattern	Fix a wrong pattern — your correction overrides with maximum confidence

New CLI Commands

slm useful 42 87           # Mark memories as useful (ranking feedback)
slm patterns list           # See learned tech preferences
slm learning status         # Learning system diagnostics
slm learning reset          # Delete all behavioral data (memories preserved)
slm engagement              # Local engagement health metrics

Upgrade: npm install -g superlocalmemory@latest — Learning dependencies install automatically.

Learning System Guide → | Upgrade Guide → | Full Changelog

---

Previous: v2.6.5 — Interactive Knowledge Graph

• Fully interactive visualization with zoom, pan, click-to-explore (Cytoscape.js)
• 6 layout algorithms, smart cluster filtering, 10,000+ node performance
• Mobile & accessibility support: touch gestures, keyboard nav, screen reader

Previous: v2.6 — Security & Scale

What's New in v2.6

SuperLocalMemory is now production-hardened with security, performance, and scale improvements:

• Trust Enforcement — Bayesian scoring actively protects your memory. Agents with trust below 0.3 are blocked from write/delete operations.
• Profile Isolation — Memory profiles fully sandboxed. Zero cross-profile data leakage.
• Rate Limiting — Protects against memory flooding from misbehaving agents.
• HNSW-Accelerated Graphs — Knowledge graph edge building uses HNSW index for faster construction at scale.
• Hybrid Search Engine — Combined semantic + FTS5 + graph retrieval for maximum accuracy.

v2.5 highlights (included): Real-time event stream, WAL-mode concurrent writes, agent tracking, memory provenance, 28 API endpoints.

Upgrade: npm install -g superlocalmemory@latest

Interactive Architecture Diagram | Architecture Doc | Full Changelog

---

The Problem

Every time you start a new Claude session:

You: "Remember that authentication bug we fixed last week?"
Claude: "I don't have access to previous conversations..."
You: *sighs and explains everything again*

AI assistants forget everything between sessions. You waste time re-explaining your:

• Project architecture
• Coding preferences
• Previous decisions
• Debugging history

---

The Solution

# Install in one command
npm install -g superlocalmemory

# Save a memory
superlocalmemoryv2:remember "Fixed auth bug - JWT tokens were expiring too fast, increased to 24h"

# Later, in a new session...
superlocalmemoryv2:recall "auth bug"
# ✓ Found: "Fixed auth bug - JWT tokens were expiring too fast, increased to 24h"

Your AI now remembers everything. Forever. Locally. For free.

---

🚀 Quick Start

Install (One Command)

npm install -g superlocalmemory

Or clone manually:

git clone https://github.com/varun369/SuperLocalMemoryV2.git && cd SuperLocalMemoryV2 && ./install.sh

Both methods auto-detect and configure 17+ IDEs and AI tools — Cursor, VS Code/Copilot, Codex, Claude, Windsurf, Gemini CLI, JetBrains, and more.

Verify Installation

superlocalmemoryv2:status
# ✓ Database: OK (0 memories)
# ✓ Graph: Ready
# ✓ Patterns: Ready

That's it. No Docker. No API keys. No cloud accounts. No configuration.

Launch Dashboard

# Start the interactive web UI
python3 ~/.claude-memory/ui_server.py

# Opens at http://localhost:8765
# Features: Timeline, search, interactive graph, statistics

---

💡 Why SuperLocalMemory?

For Developers Who Use AI Daily

Scenario	Without Memory	With SuperLocalMemory
New Claude session	Re-explain entire project	recall "project context" → instant context
Debugging	"We tried X last week..." starts over	Knowledge graph shows related past fixes
Code preferences	"I prefer React..." every time	Pattern learning knows your style
Multi-project	Context constantly bleeds	Separate profiles per project

Built on 2026 Research

Not another simple key-value store. SuperLocalMemory implements cutting-edge memory architecture backed by peer-reviewed research:

• PageIndex (Meta AI) → Hierarchical memory organization
• GraphRAG (Microsoft) → Knowledge graph with auto-clustering
• xMemory (Stanford) → Identity pattern learning
• A-RAG → Multi-level retrieval with context awareness
• LambdaRank (Burges 2010, MO-LightGBM SIGIR 2025) → Adaptive re-ranking (v2.7)
• TSW-PrefixSpan (IEEE 2020) → Time-weighted workflow pattern mining (v2.7)
• MACLA (arXiv:2512.18950) → Bayesian temporal confidence scoring (v2.7)
• FCS (LREC 2024) → Hierarchical cold-start mitigation (v2.7)

The only open-source implementation combining all eight approaches — entirely locally.

See research citations →

---

✨ Features

Multi-Layer Memory Architecture

View Interactive Architecture Diagram — Click any layer for details, research references, and file paths.

┌─────────────────────────────────────────────────────────────┐
│  Layer 9: VISUALIZATION (v2.2+)                             │
│  Interactive dashboard: timeline, graph explorer, analytics │
├─────────────────────────────────────────────────────────────┤
│  Layer 8: HYBRID SEARCH (v2.2+)                             │
│  Combines: Semantic + FTS5 + Graph traversal                │
├─────────────────────────────────────────────────────────────┤
│  Layer 7: UNIVERSAL ACCESS                                  │
│  MCP + Skills + CLI (works everywhere)                      │
│  17+ IDEs with single database                              │
├─────────────────────────────────────────────────────────────┤
│  Layer 6: MCP INTEGRATION                                   │
│  Model Context Protocol: 12 tools, 6 resources, 2 prompts   │
│  Auto-configured for Cursor, Windsurf, Claude               │
├─────────────────────────────────────────────────────────────┤
│  Layer 5½: ADAPTIVE LEARNING (v2.7 — NEW)                   │
│  Three-layer learning: tech prefs + project context + flow  │
│  LightGBM LambdaRank re-ranking (fully local, no cloud)    │
│  Research: eKNOW 2025, MACLA, TSW-PrefixSpan, LREC 2024    │
├─────────────────────────────────────────────────────────────┤
│  Layer 5: SKILLS LAYER                                      │
│  7 universal slash-commands for AI assistants               │
│  Compatible with Claude Code, Continue, Cody                │
├─────────────────────────────────────────────────────────────┤
│  Layer 4: PATTERN LEARNING + MACLA                          │
│  Bayesian confidence scoring (arXiv:2512.18950)             │
│  "You prefer React over Vue" (73% confidence)               │
├─────────────────────────────────────────────────────────────┤
│  Layer 3: KNOWLEDGE GRAPH + HIERARCHICAL CLUSTERING         │
│  Recursive Leiden algorithm: "Python" → "FastAPI" → "Auth" │
│  Community summaries with TF-IDF structured reports         │
├─────────────────────────────────────────────────────────────┤
│  Layer 2: HIERARCHICAL INDEX                                │
│  Tree structure for fast navigation                         │
│  O(log n) lookups instead of O(n) scans                     │
├─────────────────────────────────────────────────────────────┤
│  Layer 1: RAW STORAGE                                       │
│  SQLite + Full-text search + TF-IDF vectors                 │
│  Compression: 60-96% space savings                          │
└─────────────────────────────────────────────────────────────┘

Key Capabilities

• Adaptive Learning System — Learns your tech preferences, workflow patterns, and project context. Personalizes recall ranking using local ML (LightGBM). Zero cloud dependency. New in v2.7
• Knowledge Graphs — Automatic relationship discovery. Interactive visualization with zoom, pan, click.
• Pattern Learning — Learns your coding preferences and style automatically.
• Multi-Profile Support — Isolated contexts for work, personal, clients. Zero context bleeding.
• Hybrid Search — Semantic + FTS5 + Graph retrieval combined for maximum accuracy.
• Visualization Dashboard — Web UI for timeline, search, graph exploration, analytics.
• Framework Integrations — Use with LangChain and LlamaIndex applications.
• Real-Time Events — Live notifications via SSE/WebSocket/Webhooks when memories change.

---

🌐 Works Everywhere

SuperLocalMemory V2 is the ONLY memory system that works across ALL your tools:

Supported IDEs & Tools

Tool	Integration	How It Works
Claude Code	✅ Skills + MCP	/superlocalmemoryv2:remember
Cursor	✅ MCP + Skills	AI uses memory tools natively
Windsurf	✅ MCP + Skills	Native memory access
Claude Desktop	✅ MCP	Built-in support
OpenAI Codex	✅ MCP + Skills	Auto-configured (TOML)
VS Code / Copilot	✅ MCP + Skills	.vscode/mcp.json
Continue.dev	✅ MCP + Skills	/slm-remember
Cody	✅ Custom Commands	/slm-remember
Gemini CLI	✅ MCP + Skills	Native MCP + skills
JetBrains IDEs	✅ MCP	Via AI Assistant settings
Zed Editor	✅ MCP	Native MCP tools
Aider	✅ Smart Wrapper	aider-smart with context
Any Terminal	✅ Universal CLI	slm remember "content"

Three Ways to Access

1. MCP (Model Context Protocol) — Auto-configured for Cursor, Windsurf, Claude Desktop

   • AI assistants get natural access to your memory
   • No manual commands needed
   • "Remember that we use FastAPI" just works

2. Skills & Commands — For Claude Code, Continue.dev, Cody

   • /superlocalmemoryv2:remember in Claude Code
   • /slm-remember in Continue.dev and Cody
   • Familiar slash command interface

3. Universal CLI — Works in any terminal or script

   • slm remember "content" - Simple, clean syntax
   • slm recall "query" - Search from anywhere
   • aider-smart - Aider with auto-context injection

All three methods use the SAME local database. No data duplication, no conflicts.

Complete setup guide for all tools →

---

🆚 vs Alternatives

The Hard Truth About "Free" Tiers

Solution	Free Tier Limits	Paid Price	What's Missing
Mem0	10K memories, limited API	Usage-based	No pattern learning, not local
Zep	Limited credits	$50/month	Credit system, cloud-only
Supermemory	1M tokens, 10K queries	$19-399/mo	Not local, no graphs
Personal.AI	❌ No free tier	$33/month	Cloud-only, closed ecosystem
Letta/MemGPT	Self-hosted (complex)	TBD	Requires significant setup
SuperLocalMemory V2	Unlimited	$0 forever	Nothing.

What Actually Matters

Feature	Mem0	Zep	Khoj	Letta	SuperLocalMemory V2
Works in Cursor	Cloud Only	❌	❌	❌	✅ Local
Works in Windsurf	Cloud Only	❌	❌	❌	✅ Local
Works in VS Code	3rd Party	❌	Partial	❌	✅ Native
Universal CLI	❌	❌	❌	❌	✅
Multi-Layer Architecture	❌	❌	❌	❌	✅
Pattern Learning	❌	❌	❌	❌	✅
Adaptive ML Ranking	Cloud LLM	❌	❌	❌	✅ Local ML
Knowledge Graphs	✅	✅	❌	❌	✅
100% Local	❌	❌	Partial	Partial	✅
GDPR by Design	❌	❌	❌	❌	✅
Zero Setup	❌	❌	❌	❌	✅
Completely Free	Limited	Limited	Partial	✅	✅

SuperLocalMemory V2 is the ONLY solution that:

• ✅ Learns and adapts locally — no cloud LLM needed for personalization
• ✅ Works across 17+ IDEs and CLI tools
• ✅ Remains 100% local (no cloud dependencies)
• ✅ GDPR Article 17 compliant — one-command data erasure
• ✅ Completely free with unlimited memories

See full competitive analysis →

---

⚡ Measured Performance

All numbers measured on real hardware (Apple M4 Pro, 24GB RAM). No estimates — real benchmarks.

Search Speed

Database Size	Median Latency	P95 Latency
100 memories	10.6ms	14.9ms
500 memories	65.2ms	101.7ms
1,000 memories	124.3ms	190.1ms

For typical personal use (under 500 memories), search results return faster than you blink.

Concurrent Writes — Zero Errors

Scenario	Writes/sec	Errors
1 AI tool writing	204/sec	0
2 AI tools simultaneously	220/sec	0
5 AI tools simultaneously	130/sec	0

WAL mode + serialized write queue = zero "database is locked" errors, ever.

Storage

10,000 memories = 13.6 MB on disk (~1.4 KB per memory). Your entire AI memory history takes less space than a photo.

Graph Construction

Memories	Build Time
100	0.28s
1,000	10.6s

Leiden clustering discovers 6-7 natural topic communities automatically.

Full benchmark details →

---

🔧 CLI Commands

# Memory Operations
superlocalmemoryv2:remember "content" --tags tag1,tag2  # Save memory
superlocalmemoryv2:recall "search query"                 # Search
superlocalmemoryv2:list                                  # Recent memories
superlocalmemoryv2:status                                # System health

# Profile Management
superlocalmemoryv2:profile list                          # Show all profiles
superlocalmemoryv2:profile create <name>                 # New profile
superlocalmemoryv2:profile switch <name>                 # Switch context

# Knowledge Graph
python ~/.claude-memory/graph_engine.py build            # Build graph
python ~/.claude-memory/graph_engine.py stats            # View clusters

# Pattern Learning
python ~/.claude-memory/pattern_learner.py update        # Learn patterns
python ~/.claude-memory/pattern_learner.py context 0.5   # Get identity

# Visualization Dashboard
python ~/.claude-memory/ui_server.py                     # Launch web UI

Complete CLI reference →

---

📖 Documentation

Guide	Description
Quick Start	Get running in 5 minutes
Installation	Detailed setup instructions
Visualization Dashboard	Interactive web UI guide
Interactive Graph	Graph exploration guide (NEW v2.6.5)
Framework Integrations	LangChain & LlamaIndex setup
Knowledge Graph	How clustering works
Pattern Learning	Identity extraction
API Reference	Python API documentation

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Areas for contribution:

• Additional pattern categories
• Performance optimizations
• Integration with more AI assistants
• Documentation improvements

---

💖 Support This Project

If SuperLocalMemory saves you time, consider supporting its development:

• ⭐ Star this repo — helps others discover it
• 🐛 Report bugs — open an issue
• 💡 Suggest features — start a discussion
• ☕ Buy me a coffee — buymeacoffee.com/varunpratah
• 💸 PayPal — paypal.me/varunpratapbhardwaj
• 💖 Sponsor — GitHub Sponsors

---

📜 License

MIT License — use freely, even commercially. Just include the license.

---

👨‍💻 Author

Varun Pratap Bhardwaj — Solution Architect

Building tools that make AI actually useful for developers.

---

100% local. 100% private. 100% yours.