🏗️ Infrastructure Guide

Overview

Eliza's infrastructure is built on a flexible database architecture that supports multiple adapters and efficient data storage mechanisms for AI agent interactions, memory management, and relationship tracking.

Core Components

Database Adapters

Eliza supports multiple database backends through a pluggable adapter system:

• PostgreSQL - Full-featured adapter with vector search capabilities
• SQLite - Lightweight local database option
• SQL.js - In-memory database for testing and development
• Supabase - Cloud-hosted PostgreSQL with additional features

Schema Structure

The database schema includes several key tables:

- accounts: User and agent identities
- rooms: Conversation spaces
- memories: Vector-indexed message storage
- goals: Agent objectives and progress
- participants: Room membership tracking
- relationships: Inter-agent connections

Setting Up Infrastructure

PostgreSQL Setup

1. Install PostgreSQL Extensions

CREATE EXTENSION IF NOT EXISTS vector;
CREATE EXTENSION IF NOT EXISTS fuzzystrmatch;

2. Initialize Core Tables

-- Create base tables
CREATE TABLE accounts (
    "id" UUID PRIMARY KEY,
    "createdAt" TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    "name" TEXT,
    "username" TEXT,
    "email" TEXT NOT NULL,
    "avatarUrl" TEXT,
    "details" JSONB DEFAULT '{}'::jsonb
);

CREATE TABLE rooms (
    "id" UUID PRIMARY KEY,
    "createdAt" TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP
);

CREATE TABLE memories (
    "id" UUID PRIMARY KEY,
    "type" TEXT NOT NULL,
    "createdAt" TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    "content" JSONB NOT NULL,
    "embedding" vector(1536),
    "userId" UUID REFERENCES accounts("id"),
    "agentId" UUID REFERENCES accounts("id"),
    "roomId" UUID REFERENCES rooms("id"),
    "unique" BOOLEAN DEFAULT true NOT NULL
);

3. Set Up Indexes

CREATE INDEX idx_memories_embedding ON memories
    USING hnsw ("embedding" vector_cosine_ops);
CREATE INDEX idx_memories_type_room ON memories("type", "roomId");
CREATE INDEX idx_participants_user ON participants("userId");
CREATE INDEX idx_participants_room ON participants("roomId");

Connection Configuration

// PostgreSQL Configuration
const postgresConfig = {
  max: 20,
  idleTimeoutMillis: 30000,
  connectionTimeoutMillis: 2000,
};

// Supabase Configuration
const supabaseConfig = {
  supabaseUrl: process.env.SUPABASE_URL,
  supabaseKey: process.env.SUPABASE_KEY,
};

Memory Management

Vector Storage

The memory system uses vector embeddings for semantic search:

async function storeMemory(runtime: IAgentRuntime, content: string) {
  const embedding = await runtime.embed(content);

  await runtime.databaseAdapter.createMemory({
    type: "message",
    content: { text: content },
    embedding,
    roomId: roomId,
    userId: userId,
  });
}

Memory Retrieval

async function searchMemories(runtime: IAgentRuntime, query: string) {
  const embedding = await runtime.embed(query);

  return runtime.databaseAdapter.searchMemoriesByEmbedding(embedding, {
    match_threshold: 0.8,
    count: 10,
    tableName: "memories",
  });
}

Scaling Considerations

Database Optimization

1. Index Management

   • Use HNSW indexes for vector similarity search
   • Create appropriate indexes for frequent query patterns
   • Regularly analyze and update index statistics

2. Connection Pooling

   const pool = new Pool({
     max: 20, // Maximum pool size
     idleTimeoutMillis: 30000,
     connectionTimeoutMillis: 2000,
   });

3. Query Optimization

   • Use prepared statements
   • Implement efficient pagination
   • Optimize vector similarity searches

High Availability

1. Database Replication

   • Set up read replicas for scaling read operations
   • Configure streaming replication for failover
   • Implement connection retry logic

2. Backup Strategy

   -- Regular backups
   pg_dump -Fc mydb > backup.dump
   
   -- Point-in-time recovery
   pg_basebackup -D backup -Fp -Xs -P

Security

Access Control

1. Row Level Security

ALTER TABLE memories ENABLE ROW LEVEL SECURITY;

CREATE POLICY "memories_isolation" ON memories
    USING (auth.uid() = "userId" OR auth.uid() = "agentId");

2. Role Management

-- Create application role
CREATE ROLE app_user;

-- Grant necessary permissions
GRANT SELECT, INSERT ON memories TO app_user;
GRANT USAGE ON SCHEMA public TO app_user;

Data Protection

1. Encryption

   • Use TLS for connections
   • Encrypt sensitive data at rest
   • Implement key rotation

2. Audit Logging

CREATE TABLE logs (
    "id" UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    "createdAt" TIMESTAMPTZ DEFAULT CURRENT_TIMESTAMP,
    "userId" UUID NOT NULL REFERENCES accounts("id"),
    "body" JSONB NOT NULL,
    "type" TEXT NOT NULL,
    "roomId" UUID NOT NULL REFERENCES rooms("id")
);

Monitoring

Health Checks

async function checkDatabaseHealth(): Promise<boolean> {
  try {
    await db.query("SELECT 1");
    return true;
  } catch (error) {
    console.error("Database health check failed:", error);
    return false;
  }
}

Performance Metrics

Track key metrics:

• Query performance
• Connection pool utilization
• Memory usage
• Vector search latency

Maintenance

Regular Tasks

1. Vacuum Operations

-- Regular vacuum
VACUUM ANALYZE memories;

-- Analyze statistics
ANALYZE memories;

2. Index Maintenance

-- Reindex vector similarity index
REINDEX INDEX idx_memories_embedding;

Data Lifecycle

1. Archival Strategy

   • Archive old conversations
   • Compress inactive memories
   • Implement data retention policies

2. Cleanup Jobs

async function cleanupOldMemories() {
  const cutoffDate = new Date();
  cutoffDate.setMonth(cutoffDate.getMonth() - 6);

  await db.query(
    `
        DELETE FROM memories 
        WHERE "createdAt" < $1
    `,
    [cutoffDate],
  );
}

Troubleshooting

Common Issues

1. Connection Problems

   • Check connection pool settings
   • Verify network connectivity
   • Review firewall rules

2. Performance Issues

   • Analyze query plans
   • Check index usage
   • Monitor resource utilization

3. Vector Search Problems

   • Verify embedding dimensions
   • Check similarity thresholds
   • Review index configuration

Diagnostic Queries

-- Check connection status
SELECT * FROM pg_stat_activity;

-- Analyze query performance
EXPLAIN ANALYZE
SELECT * FROM memories
WHERE embedding <-> $1 < 0.3
LIMIT 10;

-- Monitor index usage
SELECT schemaname, tablename, indexname, idx_scan
FROM pg_stat_user_indexes;

Further Reading

• PostgreSQL Documentation