Features

SuperLocalMemory V2 combines cutting-edge 2026 research with practical, production-ready features.

🏗️ 10-Layer Universal Architecture

The only memory system with both MCP (agent-to-tool) and A2A (agent-to-agent) protocol support. Each layer enhances the previous one.

┌─────────────────────────────────────────────────────┐
│ Layer 10: A2A AGENT COLLABORATION (v2.5.0)         │
│  Agent-to-Agent protocol for multi-agent sync      │
├─────────────────────────────────────────────────────┤
│ Layer 9: VISUALIZATION (v2.2.0)                    │
│  Interactive dashboard, timeline, search explorer   │
├─────────────────────────────────────────────────────┤
│ Layer 8: HYBRID SEARCH (v2.2.0)                    │
│  Semantic + FTS5 + Graph = 80ms with max accuracy  │
├─────────────────────────────────────────────────────┤
│ Layer 7: UNIVERSAL ACCESS                          │
│  MCP + Skills + CLI (works everywhere)             │
├─────────────────────────────────────────────────────┤
│ Layer 6: MCP INTEGRATION                           │
│  6 tools, 4 resources, 2 prompts                   │
├─────────────────────────────────────────────────────┤
│ Layer 5: SKILLS LAYER                              │
│  6 universal slash-commands                        │
├─────────────────────────────────────────────────────┤
│ Layer 4: PATTERN LEARNING + MACLA (v2.4.0)         │
│  Bayesian Beta-Binomial confidence scoring         │
├─────────────────────────────────────────────────────┤
│ Layer 3: KNOWLEDGE GRAPH + HIERARCHICAL (v2.4.1)   │
│  Recursive Leiden clustering + community summaries │
├─────────────────────────────────────────────────────┤
│ Layer 2: HIERARCHICAL INDEX                        │
│  Tree structure for O(log n) navigation            │
├─────────────────────────────────────────────────────┤
│ Layer 1: RAW STORAGE                               │
│  SQLite + FTS5 + TF-IDF vectors                    │
└─────────────────────────────────────────────────────┘

🤝 A2A Agent Collaboration (Coming v2.5.0)

NEW

Multi-agent collaboration through the A2A Protocol (Google/Linux Foundation). Your AI tools finally talk to each other through shared memory.

Agent Discovery - AI agents discover SuperLocalMemory via Agent Cards
Real-Time Sync - Cursor saves a preference, Claude Desktop knows instantly
Secure by Default - Per-agent permissions, local-only auth, audit logging
Dual Protocol - MCP for tool access + A2A for agent collaboration

Why both MCP + A2A? MCP connects tools to memory (vertical). A2A connects agents to each other through memory (horizontal). Together, they make SuperLocalMemory the complete memory backbone for multi-agent workflows.

Core Features

🔍 Hybrid Search (NEW v2.2.0)

Combines three search strategies for maximum accuracy:

Semantic Search - TF-IDF + cosine similarity for conceptual queries
Full-Text Search - SQLite FTS5 with ranking for exact phrases
Graph-Enhanced - Knowledge graph traversal for related concepts
Hybrid Mode - All three combined for optimal results

Measured: 10.6ms median search latency for typical databases. MRR 0.90 — first relevant result at position 1 for 9/10 queries.

📊 Visualization Dashboard (NEW v2.2.0)

Interactive web-based UI for exploring memories visually:

Timeline View - See memories chronologically with importance indicators
Search Explorer - Real-time semantic search with score visualization
Graph Visualization - Interactive knowledge graph with clusters
Statistics Dashboard - Memory trends, tag clouds, pattern insights

python ~/.claude-memory/ui_server.py
# Opens at http://localhost:8765

🕸️ Knowledge Graph

Auto-discovers relationships you didn't know existed:

TF-IDF Entity Extraction - Identifies important terms and concepts
Leiden Clustering - Groups related memories into communities
Hierarchical Leiden (v2.4.1) - Recursive sub-clustering: "Python" → "FastAPI" → "Auth patterns"
Community Summaries (v2.4.1) - TF-IDF structured reports: key topics, projects, categories per cluster
Auto-naming - Clusters get descriptive names ("Auth & Tokens", "Performance")
Relationship Discovery - Finds connections between memories

Example: Search "authentication" and get JWT, OAuth, session management—even if never tagged together.

🎯 Pattern Learning + MACLA Confidence (v2.4.0)

Learns your coding identity with research-grounded Bayesian scoring (arXiv:2512.18950):

Framework Preferences - "You prefer React over Vue" (73% confidence)
Coding Style - "Performance over readability" (58% confidence)
Testing Approach - "Jest + React Testing Library" (65% confidence)
API Style - "REST over GraphQL" (81% confidence)

python ~/.claude-memory/pattern_learner.py update
python ~/.claude-memory/pattern_learner.py context 0.5

🌐 Universal Integration

Three ways to access the same database:

MCP Protocol

Real-time, IDE-integrated. Auto-configured for Cursor, Windsurf, Claude Desktop, Continue, and more.

Universal Skills

Slash commands for Claude Code, Continue.dev, Cody, Cursor, Windsurf.

CLI

Works in any terminal, scripts, automation. Simple slm command.

👤 Multi-Profile Support

Completely isolated contexts for different projects:

Work Profile - Day job memories and patterns
Personal Profile - Side projects and learning
Client Profiles - Separate for each client (no context bleeding)

slm switch-profile work
slm switch-profile personal
slm switch-profile client-acme

🔄 Auto-Backup System (NEW v2.4.0)

Never lose your memories. Automatic SQLite backups with full control:

Configurable Intervals - Daily, weekly, or custom schedule
Retention Policies - Keep N most recent backups automatically
One-Click Restore - Restore from any backup via dashboard or CLI
Safe Concurrent Backup - Uses SQLite's backup() API, no locking

python ~/.claude-memory/auto_backup.py backup   # Manual backup
python ~/.claude-memory/auto_backup.py list     # List backups
python ~/.claude-memory/auto_backup.py status   # Backup status

💾 Progressive Compression

3-tier storage system saves 60-96% space:

Tier	Age	Compression	Savings
Tier 1	0-30 days	None	Active memories
Tier 2	30-90 days	Progressive summarization	60% savings
Tier 3	90+ days	JSON archival	96% savings

⚡ Measured Performance

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

10.6ms

Search Latency (100 memories)

220/sec

Concurrent Writes

1.9 KB

Per Memory at Scale

0 Errors

10 Concurrent Agents

1.0

Trust Gap (perfect separation)

13.6 MB

10,000 Memories on Disk

0.28s

Graph Build (100 memories)

0.90

MRR (Search Quality)

LoCoMo benchmark results coming soon — evaluation against the standardized LoCoMo long-conversation memory benchmark (Snap Research, ACL 2024).
View full benchmark details →

🔬 Built on Research

📄 PageIndex (VectifyAI)

Hierarchical memory organization with tree structures for O(log n) navigation.

🕸️ GraphRAG (Microsoft)

Knowledge graph with Leiden clustering for community detection.

🎯 MemoryBank (AAAI 2024)

Identity pattern learning that understands your preferences.

⚡ A-RAG

Multi-level retrieval with context awareness.

📸 See It In Action

Visual tour of SuperLocalMemory V2's dashboard and capabilities.

Experience These Features

Install SuperLocalMemory V2 in 5 minutes. All features included, free forever.

⭐ Star on GitHub & Get Started