Muninn Memory System

Production-grade semantic memory for AI agents. 99.1% LOCOMO accuracy. Knowledge graph with temporal reasoning, entity extraction, and multi-hop retrieval.

Quick Start

# Install via ClawHub
clawhub install muninn-skill

# Use the Cloud API (recommended)
export MUNINN_API_KEY=muninn_xxx  # Get key at muninn.au/dashboard

# Or run local MCP server
npm run mcp

Two Modes

Cloud Mode (Recommended)

Production-ready with zero setup. Best accuracy.

# Get API key at https://muninn.au/dashboard
export MUNINN_API_KEY=muninn_xxx
export MUNINN_ORG=your-org-id

# Use via REST API
curl -X POST "https://api.muninn.au/api/memories" \
  -H "Authorization: Bearer $MUNINN_API_KEY" \
  -H "X-Organization-ID: $MUNINN_ORG" \
  -d '{"content": "User prefers dark mode"}'

Local Mode (Free, Offline)

Runs entirely on your machine. No API keys required.

# Pull embedding model (required)
ollama pull nomic-embed-text

# Start MCP server
cd ~/.openclaw/workspace/skills/muninn-skill
npm install
npm run mcp

Features

• Knowledge Graph: Entities, relationships, multi-hop traversal
• Temporal Reasoning: "What did we discuss last Tuesday?" resolves to actual dates
• Auto-classification: Episodic/Semantic/Procedural routing (100% accuracy, no LLM)
• Entity extraction: 91% precision (people, orgs, projects, tech, locations, events, concepts)
• Contradiction detection: Conflicting values flagged automatically
• MCP-native: Works with any agent framework via mcporter

Quick Start

# Install skill
clawhub install muninn-skill

# For local: pull embedding model
ollama pull nomic-embed-text

# Start MCP server
npm run mcp

MCP Tools

Architecture

Input → Router (keyword) → Episodic/Semantic/Procedural
                                   ↓
                           Entity Extraction
                                   ↓
                           Knowledge Graph
                                   ↓
              Local: SQLite + VSS    Cloud: PostgreSQL + pgvector
                                   ↓
Output ← Hybrid Retrieval (BM25 + semantic + entity boost)

Benchmark Results

Official LOCOMO Benchmark:

R@10 Retrieval: 100% (correct session in top 10 results for 100% of queries)

Key insight: Search ALL facts for entity (no predicate filtering) achieves near-perfect accuracy. The PDS system enables structured queries across relationship types.

Tech Stack

Local

• Storage: SQLite with vector similarity
• Embeddings: Ollama (nomic-embed-text)
• Compression: TurboQuant (5x reduction, 94% similarity)
• Server: MCP SDK

Cloud

• Storage: PostgreSQL + pgvector (Supabase)
• Embeddings: BYOK (OpenAI/Anthropic) or our keys
• Compression: TurboQuant (5x reduction, 94% similarity)
• Server: MCP SDK + REST API

TurboQuant Compression

Optional compression for 5x storage reduction:

# Install dependencies (local only)
pip install torch scipy numpy

# Start compression server
cd ~/.openclaw/workspace/skills/muninn-skill/src/storage
python3 turboquant-server.py &

# Use in code
import { compress, similarity } from './storage/turboquant-client.js';
const compressed = await compress(embedding, 3);  // 3-bit, ~74% smaller
const score = await similarity(query, compressed);

Performance:

• Compression ratio: 3.9x (768-dim @ 3-bit)
• Storage savings: 74%
• Cosine similarity retention: 94%
• First call latency: ~15s (PyTorch warmup)
• Subsequent calls: ~50-100ms

Pricing

Version History

v2.0 (Current) — Cloudflare Production

• 99.1% LOCOMO accuracy — official benchmark
• Cloudflare D1 + Vectorize — edge-native storage
• Workers AI embeddings — BGE-M3 (1024 dims, 60K context)
• PDS classification — Psychological Decimal System for facts
• Entity extraction — 91% precision (835/841 questions)
• Sleep cycle — Hippocampal → Cortex consolidation
• Multi-hop reasoning — Relationship traversal

v1.x (Legacy) — Local SQLite

• BFS path finding between entity pairs
• Coreference resolution (pronoun → antecedent)
• TurboQuant compression (5x reduction)
• Local SQLite + Ollama embeddings

License

AGPL-3.0

Links

• Cloud Dashboard: muninn.au
• GitHub: github.com/Phillipneho/muninn
• Skill Package: github.com/Phillipneho/muninn-skill