Custom Vector Memory

This guide shows you how to create custom vector memory implementations by extending BaseVectorMemory and implementing the IVectorMemory interface. Custom vector memories allow you to integrate with any vector database or implement specialized semantic search behaviors.

🏗️ Custom Vector Memory Architecture

🎯 When to Build Custom Vector Memory

Consider building a custom vector memory when:

Integrating New Vector Databases: Your organization uses a vector database not natively supported (e.g., Elasticsearch, MongoDB Atlas Vector Search, Redis Vector)
Custom Embedding Logic: You need specialized embedding generation (e.g., custom models, pre-processing, caching)
Specialized Search: You require advanced filtering, hybrid search, or custom ranking algorithms
Performance Optimization: You need specific optimizations for your use case (e.g., approximate nearest neighbor tuning)
Multi-Collection Management: You need to search across multiple collections with custom merging logic
Access Control: You require row-level security or tenant isolation in vector search

📚 Understanding BaseVectorMemory

The BaseVectorMemory class provides most of the functionality you need:

What BaseVectorMemory Provides

// Automatic handling of:
- Message storage and retrieval
- Embedding generation via configured provider
- Basic configuration management
- Message counting and clearing
- Export/import functionality (partial)
- System message handling

What You Need to Implement

When extending BaseVectorMemory, you must implement these key methods:

/**
 * Store a message with its vector representation
 */
function add( required any message )

/**
 * Retrieve semantically relevant messages
 */
function getRelevant( required string query, numeric limit = 5 )

/**
 * Get all stored messages (for non-vector operations)
 */
function getAll()

/**
 * Remove all messages from vector storage
 */
function clear()

/**
 * Count total messages in vector storage
 */
function count()

Key Properties in BaseVectorMemory

variables.collection           // Collection/index name
variables.embeddingProvider    // AI provider for embeddings (openai, etc.)
variables.embeddingModel       // Model to use (text-embedding-3-small, etc.)
variables.dimensions           // Vector dimensions (1536, 768, etc.)
variables.metric               // Distance metric (cosine, euclidean, dot)
variables.key                  // Conversation/session identifier

🔌 IVectorMemory Interface

The complete interface you must implement:

interface {
    /**
     * Configure the vector memory
     */
    function configure( required struct config );

    /**
     * Set the conversation key
     */
    function key( required string key );

    /**
     * Add a message to vector storage
     */
    function add( required any message );

    /**
     * Get semantically relevant messages
     */
    function getRelevant( required string query, numeric limit = 5 );

    /**
     * Get all stored messages
     */
    function getAll();

    /**
     * Clear all messages
     */
    function clear();

    /**
     * Count stored messages
     */
    function count();

    /**
     * Set system message
     */
    function setSystemMessage( required string message );

    /**
     * Get system message
     */
    function getSystemMessage();

    /**
     * Export memory state
     */
    function export();

    /**
     * Import memory state
     */
    function import( required struct data );
}

Example 1: ElasticsearchVectorMemory

A complete implementation using Elasticsearch with vector similarity search:

import bxModules.bxai.models.util.TextChunker;

/**
 * Elasticsearch Vector Memory
 * Uses Elasticsearch dense_vector field type for semantic search
 */
class extends="BaseVectorMemory" implements="IVectorMemory" {

    property name="esClient" type="any";
    property name="indexName" type="string";

    /**
     * Configure Elasticsearch connection
     */
    function configure( required struct config ) {
        // Call parent configuration
        super.configure( arguments.config );

        // Elasticsearch-specific configuration
        variables.esHost = arguments.config.esHost ?: "localhost";
        variables.esPort = arguments.config.esPort ?: 9200;
        variables.esScheme = arguments.config.esScheme ?: "http";
        variables.esUser = arguments.config.esUser ?: "";
        variables.esPassword = arguments.config.esPassword ?: "";
        variables.indexName = arguments.config.collection ?: "ai_memory";

        // Initialize Elasticsearch HTTP client
        initializeClient();

        // Ensure index exists with proper mapping
        ensureIndex();

        return this;
    }

    /**
     * Initialize Elasticsearch HTTP client
     */
    private function initializeClient() {
        variables.esClient = {
            host: variables.esHost,
            port: variables.esPort,
            scheme: variables.esScheme,
            user: variables.esUser,
            password: variables.esPassword,
            baseURL: "#variables.esScheme#://#variables.esHost#:#variables.esPort#"
        };
    }

    /**
     * Ensure index exists with vector mapping
     */
    private function ensureIndex() {
        // Check if index exists
        var exists = esRequest(
            method: "HEAD",
            endpoint: "/#variables.indexName#"
        ).statusCode == 200;

        if ( !exists ) {
            // Create index with vector field mapping
            esRequest(
                method: "PUT",
                endpoint: "/#variables.indexName#",
                body: {
                    "mappings": {
                        "properties": {
                            "id": { "type": "keyword" },
                            "key": { "type": "keyword" },
                            "role": { "type": "keyword" },
                            "content": { "type": "text" },
                            "embedding": {
                                "type": "dense_vector",
                                "dims": variables.dimensions,
                                "index": true,
                                "similarity": mapMetricToES( variables.metric )
                            },
                            "timestamp": { "type": "date" },
                            "metadata": { "type": "object", "enabled": false }
                        }
                    }
                }
            );
        }
    }

    /**
     * Add message with vector embedding
     */
    function add( required any message ) {
        // Normalize message to struct
        var msg = isSimpleValue( arguments.message )
            ? { role: "user", content: arguments.message }
            : arguments.message;

        // Generate embedding for content
        var embedding = generateEmbedding( msg.content );

        // Create document
        var doc = {
            "id": createUUID(),
            "key": variables.key,
            "role": msg.role,
            "content": msg.content,
            "embedding": embedding,
            "timestamp": now(),
            "metadata": msg.metadata ?: {}
        };

        // Index document
        esRequest(
            method: "POST",
            endpoint: "/#variables.indexName#/_doc",
            body: doc
        );

        return this;
    }

    /**
     * Get semantically relevant messages using vector similarity
     */
    function getRelevant( required string query, numeric limit = 5 ) {
        // Generate embedding for query
        var queryEmbedding = generateEmbedding( arguments.query );

        // Search using k-nearest neighbors
        var response = esRequest(
            method: "POST",
            endpoint: "/#variables.indexName#/_search",
            body: {
                "query": {
                    "bool": {
                        "must": [
                            {
                                "knn": {
                                    "embedding": {
                                        "vector": queryEmbedding,
                                        "k": arguments.limit
                                    }
                                }
                            },
                            {
                                "term": { "key": variables.key }
                            }
                        ]
                    }
                },
                "size": arguments.limit,
                "_source": [ "role", "content", "timestamp", "metadata" ]
            }
        );

        // Transform results
        var results = [];
        if ( response.hits?.hits.len() ) {
            response.hits.hits.each( hit => {
                results.append({
                    role: hit._source.role,
                    content: hit._source.content,
                    score: hit._score,
                    timestamp: hit._source.timestamp,
                    metadata: hit._source.metadata
                });
            });
        }

        return results;
    }

    /**
     * Get all messages for this key
     */
    function getAll() {
        var response = esRequest(
            method: "POST",
            endpoint: "/#variables.indexName#/_search",
            body: {
                "query": {
                    "term": { "key": variables.key }
                },
                "sort": [ { "timestamp": "asc" } ],
                "size": 10000,
                "_source": [ "role", "content", "timestamp", "metadata" ]
            }
        );

        var messages = [];
        if ( response.hits?.hits.len() ) {
            response.hits.hits.each( hit => {
                messages.append({
                    role: hit._source.role,
                    content: hit._source.content,
                    timestamp: hit._source.timestamp,
                    metadata: hit._source.metadata
                });
            });
        }

        return messages;
    }

    /**
     * Clear all messages for this key
     */
    function clear() {
        esRequest(
            method: "POST",
            endpoint: "/#variables.indexName#/_delete_by_query",
            body: {
                "query": {
                    "term": { "key": variables.key }
                }
            }
        );
        return this;
    }

    /**
     * Count messages for this key
     */
    function count() {
        var response = esRequest(
            method: "GET",
            endpoint: "/#variables.indexName#/_count",
            body: {
                "query": {
                    "term": { "key": variables.key }
                }
            }
        );
        return response.count ?: 0;
    }

    /**
     * Make HTTP request to Elasticsearch
     */
    private function esRequest(
        required string method,
        required string endpoint,
        struct body = {}
    ) {
        var url = variables.esClient.baseURL & arguments.endpoint;

        var httpReq = http( url )
            .method( arguments.method )
            .header( "Content-Type", "application/json" );

        // Add authentication if configured
        if ( variables.esClient.user.len() ) {
            var auth = toBase64( "#variables.esClient.user#:#variables.esClient.password#" );
            httpReq.header( "Authorization", "Basic #auth#" );
        }

        // Add body for non-GET/HEAD requests
        if ( ![ "GET", "HEAD" ].findNoCase( arguments.method ) && !arguments.body.isEmpty() ) {
            httpReq.body( jsonSerialize( arguments.body ) );
        }

        var response = httpReq.send();

        // Parse JSON response
        return response.getStatusCode() == 200 || response.getStatusCode() == 201
            ? jsonDeserialize( response.getFileContent() )
            : {};
    }

    /**
     * Map our metric names to Elasticsearch similarity functions
     */
    private function mapMetricToES( required string metric ) {
        return {
            "cosine": "cosine",
            "euclidean": "l2_norm",
            "dot": "dot_product"
        }[ arguments.metric ] ?: "cosine";
    }

    /**
     * Generate embedding using configured provider
     */
    private function generateEmbedding( required string text ) {
        // Use aiEmbed() BIF
        var result = aiEmbed(
            input: arguments.text,
            provider: variables.embeddingProvider,
            model: variables.embeddingModel
        );

        return result.embeddings[ 1 ];
    }
}

Usage Example

// Create Elasticsearch vector memory
memory = new ElasticsearchVectorMemory().configure({
    collection: "customer_support",
    esHost: "localhost",
    esPort: 9200,
    esUser: "elastic",
    esPassword: "password",
    embeddingProvider: "openai",
    embeddingModel: "text-embedding-3-small",
    dimensions: 1536,
    metric: "cosine"
});

// Use with agent
agent = aiAgent(
    name: "SupportBot",
    memory: memory
);

Example 2: RedisVectorMemory

Implementation using Redis with RediSearch vector similarity:

/**
 * Redis Vector Memory
 * Uses Redis Stack with vector similarity search (RediSearch)
 */
class extends="BaseVectorMemory" implements="IVectorMemory" {

    /**
     * Configure Redis connection
     */
    function configure( required struct config ) {
        super.configure( arguments.config );

        variables.redisHost = arguments.config.redisHost ?: "localhost";
        variables.redisPort = arguments.config.redisPort ?: 6379;
        variables.redisPassword = arguments.config.redisPassword ?: "";
        variables.indexName = "idx:#arguments.config.collection ?: 'ai_memory'#";

        // Initialize Redis connection
        initializeRedis();

        // Create index if needed
        ensureIndex();

        return this;
    }

    /**
     * Initialize Redis connection
     */
    private function initializeRedis() {
        variables.redisURL = "redis://#variables.redisHost#:#variables.redisPort#";

        // Test connection
        var response = redisCommand( "PING" );
        if ( response != "PONG" ) {
            throw( type="RedisConnectionException", message="Failed to connect to Redis" );
        }
    }

    /**
     * Ensure search index exists
     */
    private function ensureIndex() {
        // Check if index exists
        try {
            redisCommand( "FT.INFO", [ variables.indexName ] );
        } catch ( any e ) {
            // Create index with vector field
            redisCommand( "FT.CREATE", [
                variables.indexName,
                "ON", "HASH",
                "PREFIX", "1", "msg:#variables.key#:",
                "SCHEMA",
                "role", "TAG",
                "content", "TEXT",
                "embedding", "VECTOR", "HNSW", "6",
                    "TYPE", "FLOAT32",
                    "DIM", variables.dimensions,
                    "DISTANCE_METRIC", mapMetricToRedis( variables.metric ),
                "timestamp", "NUMERIC", "SORTABLE"
            ] );
        }
    }

    /**
     * Add message with embedding
     */
    function add( required any message ) {
        var msg = isSimpleValue( arguments.message )
            ? { role: "user", content: arguments.message }
            : arguments.message;

        var embedding = generateEmbedding( msg.content );
        var msgId = "msg:#variables.key#:#createUUID()#";

        // Store as Redis hash
        redisCommand( "HSET", [
            msgId,
            "role", msg.role,
            "content", msg.content,
            "embedding", serializeVector( embedding ),
            "timestamp", getTickCount(),
            "metadata", jsonSerialize( msg.metadata ?: {} )
        ] );

        return this;
    }

    /**
     * Get relevant messages using vector search
     */
    function getRelevant( required string query, numeric limit = 5 ) {
        var queryEmbedding = generateEmbedding( arguments.query );
        var vectorBlob = serializeVector( queryEmbedding );

        // Search using vector similarity
        var results = redisCommand( "FT.SEARCH", [
            variables.indexName,
            "*=>[KNN #arguments.limit# @embedding $vector AS score]",
            "PARAMS", "2", "vector", vectorBlob,
            "SORTBY", "score",
            "DIALECT", "2",
            "RETURN", "4", "role", "content", "timestamp", "score"
        ] );

        return parseRedisResults( results );
    }

    /**
     * Get all messages
     */
    function getAll() {
        var pattern = "msg:#variables.key#:*";
        var keys = redisCommand( "KEYS", [ pattern ] );

        var messages = [];
        keys.each( key => {
            var data = redisCommand( "HGETALL", [ key ] );
            messages.append( parseHashData( data ) );
        } );

        // Sort by timestamp
        messages.sort( ( a, b ) => compare( a.timestamp, b.timestamp ) );

        return messages;
    }

    /**
     * Clear all messages
     */
    function clear() {
        var pattern = "msg:#variables.key#:*";
        var keys = redisCommand( "KEYS", [ pattern ] );

        if ( keys.len() ) {
            redisCommand( "DEL", keys );
        }

        return this;
    }

    /**
     * Count messages
     */
    function count() {
        var pattern = "msg:#variables.key#:*";
        var keys = redisCommand( "KEYS", [ pattern ] );
        return keys.len();
    }

    /**
     * Execute Redis command via HTTP API or native client
     */
    private function redisCommand( required string command, array args = [] ) {
        // Simplified: Use HTTP API or native Redis client
        // This would use actual Redis client library in production

        var cmdArray = [ arguments.command ];
        cmdArray.append( arguments.args, true );

        // Implementation would call actual Redis client
        // For example: redisClient.execute( cmdArray )

        return executeRedisCommand( cmdArray );
    }

    /**
     * Serialize vector for Redis storage
     */
    private function serializeVector( required array vector ) {
        // Convert array of floats to binary blob
        var buffer = createObject( "java", "java.nio.ByteBuffer" )
            .allocate( arguments.vector.len() * 4 );

        arguments.vector.each( val => {
            buffer.putFloat( val castAs float );
        } );

        return buffer.array();
    }

    /**
     * Map metric to Redis format
     */
    private function mapMetricToRedis( required string metric ) {
        return {
            "cosine": "COSINE",
            "euclidean": "L2",
            "dot": "IP"
        }[ arguments.metric ] ?: "COSINE";
    }

    /**
     * Parse Redis search results
     */
    private function parseRedisResults( required array results ) {
        var messages = [];
        // Redis returns: [count, key1, [field1, value1, ...], key2, ...]
        var count = results[ 1 ];

        for ( var i = 2; i <= results.len(); i += 2 ) {
            var fields = results[ i + 1 ];
            var msg = {};

            for ( var j = 1; j <= fields.len(); j += 2 ) {
                msg[ fields[ j ] ] = fields[ j + 1 ];
            }

            messages.append( msg );
        }

        return messages;
    }

    /**
     * Generate embedding
     */
    private function generateEmbedding( required string text ) {
        var result = aiEmbed(
            input: arguments.text,
            provider: variables.embeddingProvider,
            model: variables.embeddingModel
        );
        return result.embeddings[ 1 ];
    }
}

Example 3: CachedVectorMemory

A wrapper that adds caching layer to any vector memory:

/**
 * Cached Vector Memory Wrapper
 * Adds intelligent caching to reduce embedding API calls
 */
class extends="BaseVectorMemory" implements="IVectorMemory" {

    property name="wrappedMemory" type="any";
    property name="embeddingCache" type="struct";
    property name="resultCache" type="struct";

    /**
     * Configure with wrapped memory
     */
    function configure( required struct config ) {
        super.configure( arguments.config );

        // Wrapped memory instance
        if ( !arguments.config.keyExists( "wrappedMemory" ) ) {
            throw( "CachedVectorMemory requires 'wrappedMemory' in config" );
        }

        variables.wrappedMemory = arguments.config.wrappedMemory;

        // Cache configuration
        variables.embeddingCacheTTL = arguments.config.embeddingCacheTTL ?: 3600; // 1 hour
        variables.resultCacheTTL = arguments.config.resultCacheTTL ?: 300; // 5 minutes
        variables.maxCacheSize = arguments.config.maxCacheSize ?: 1000;

        // Initialize caches
        variables.embeddingCache = {};
        variables.resultCache = {};

        return this;
    }

    /**
     * Add message (cache embedding)
     */
    function add( required any message ) {
        var msg = isSimpleValue( arguments.message )
            ? { role: "user", content: arguments.message }
            : arguments.message;

        // Check if we've already embedded this exact content
        var cacheKey = hash( msg.content, "MD5" );

        if ( !variables.embeddingCache.keyExists( cacheKey ) ) {
            // Not cached, will generate embedding
            println( "Cache MISS: Generating new embedding" );
        } else {
            println( "Cache HIT: Reusing cached embedding" );
            // Could inject cached embedding into wrapped memory here
        }

        // Delegate to wrapped memory
        variables.wrappedMemory.add( arguments.message );

        // Invalidate result cache
        variables.resultCache = {};

        return this;
    }

    /**
     * Get relevant with result caching
     */
    function getRelevant( required string query, numeric limit = 5 ) {
        // Create cache key from query + limit
        var cacheKey = hash( arguments.query & arguments.limit, "MD5" );

        // Check result cache
        if ( variables.resultCache.keyExists( cacheKey ) ) {
            var cached = variables.resultCache[ cacheKey ];

            // Check if still valid
            if ( getTickCount() - cached.timestamp < variables.resultCacheTTL * 1000 ) {
                println( "Result cache HIT for query: '#left(arguments.query, 30)#...'" );
                return cached.results;
            } else {
                println( "Result cache EXPIRED" );
                variables.resultCache.delete( cacheKey );
            }
        }

        println( "Result cache MISS: Executing search" );

        // Execute search
        var results = variables.wrappedMemory.getRelevant(
            query: arguments.query,
            limit: arguments.limit
        );

        // Cache results
        variables.resultCache[ cacheKey ] = {
            results: results,
            timestamp: getTickCount()
        };

        // Enforce max cache size
        if ( variables.resultCache.count() > variables.maxCacheSize ) {
            pruneCache( variables.resultCache );
        }

        return results;
    }

    /**
     * Delegate getAll to wrapped memory
     */
    function getAll() {
        return variables.wrappedMemory.getAll();
    }

    /**
     * Clear wrapped memory and caches
     */
    function clear() {
        variables.wrappedMemory.clear();
        variables.embeddingCache = {};
        variables.resultCache = {};
        return this;
    }

    /**
     * Delegate count to wrapped memory
     */
    function count() {
        return variables.wrappedMemory.count();
    }

    /**
     * Prune cache to max size (LRU)
     */
    private function pruneCache( required struct cache ) {
        // Simple pruning: remove oldest entries
        var entries = [];

        for ( var key in arguments.cache ) {
            entries.append({
                key: key,
                timestamp: arguments.cache[ key ].timestamp
            });
        }

        // Sort by timestamp
        entries.sort( ( a, b ) => compare( a.timestamp, b.timestamp ) );

        // Remove oldest 20%
        var removeCount = ceiling( entries.len() * 0.2 );
        for ( var i = 1; i <= removeCount; i++ ) {
            arguments.cache.delete( entries[ i ].key );
        }
    }
}

Usage Example

// Wrap any vector memory with caching
baseMemory = aiMemory( "pinecone", {
    apiKey: getSystemSetting( "PINECONE_API_KEY" ),
    collection: "support_kb",
    embeddingProvider: "openai",
    embeddingModel: "text-embedding-3-small"
} );

// Add caching layer
cachedMemory = new CachedVectorMemory().configure({
    wrappedMemory: baseMemory,
    embeddingCacheTTL: 7200,  // 2 hours
    resultCacheTTL: 600,       // 10 minutes
    maxCacheSize: 500
});

// Use cached memory with agent
agent = aiAgent(
    name: "CachedBot",
    memory: cachedMemory
);

Example 4: MultiCollectionVectorMemory

Search across multiple collections with custom ranking:

/**
 * Multi-Collection Vector Memory
 * Searches multiple collections and merges results intelligently
 */
class extends="BaseVectorMemory" implements="IVectorMemory" {

    property name="collections" type="array";
    property name="collectionWeights" type="struct";

    /**
     * Configure multiple collections
     */
    function configure( required struct config ) {
        super.configure( arguments.config );

        // Multiple collection configurations
        if ( !arguments.config.keyExists( "collections" ) ) {
            throw( "MultiCollectionVectorMemory requires 'collections' array" );
        }

        variables.collections = [];
        variables.collectionWeights = arguments.config.collectionWeights ?: {};

        // Initialize each collection memory
        arguments.config.collections.each( collectionConfig => {
            var memory = aiMemory( "boxvector", collectionConfig );
            variables.collections.append({
                name: collectionConfig.collection,
                memory: memory,
                weight: variables.collectionWeights[ collectionConfig.collection ] ?: 1.0
            });
        } );

        return this;
    }

    /**
     * Add to all collections
     */
    function add( required any message ) {
        variables.collections.each( coll => {
            coll.memory.add( arguments.message );
        } );
        return this;
    }

    /**
     * Search all collections and merge results
     */
    function getRelevant( required string query, numeric limit = 5 ) {
        var allResults = [];

        // Search each collection
        variables.collections.each( coll => {
            var results = coll.memory.getRelevant(
                query: arguments.query,
                limit: arguments.limit
            );

            // Add collection metadata and apply weight
            results.each( result => {
                result.collection = coll.name;
                result.originalScore = result.score;
                result.score = result.score * coll.weight;
                allResults.append( result );
            } );
        } );

        // Sort by weighted score
        allResults.sort( ( a, b ) => {
            return a.score > b.score ? -1 : 1;
        } );

        // Return top N results
        return allResults.slice( 1, min( arguments.limit, allResults.len() ) );
    }

    /**
     * Get all from all collections
     */
    function getAll() {
        var allMessages = [];

        variables.collections.each( coll => {
            var messages = coll.memory.getAll();
            messages.each( msg => {
                msg.collection = coll.name;
                allMessages.append( msg );
            } );
        } );

        return allMessages;
    }

    /**
     * Clear all collections
     */
    function clear() {
        variables.collections.each( coll => {
            coll.memory.clear();
        } );
        return this;
    }

    /**
     * Count across all collections
     */
    function count() {
        var total = 0;
        variables.collections.each( coll => {
            total += coll.memory.count();
        } );
        return total;
    }
}

Usage Example

// Create multi-collection memory
memory = new MultiCollectionVectorMemory().configure({
    collections: [
        {
            collection: "product_docs",
            embeddingProvider: "openai",
            embeddingModel: "text-embedding-3-small",
            dimensions: 1536
        },
        {
            collection: "support_tickets",
            embeddingProvider: "openai",
            embeddingModel: "text-embedding-3-small",
            dimensions: 1536
        },
        {
            collection: "faq_database",
            embeddingProvider: "openai",
            embeddingModel: "text-embedding-3-small",
            dimensions: 1536
        }
    ],
    collectionWeights: {
        "product_docs": 1.5,      // Prioritize product docs
        "support_tickets": 1.0,   // Standard priority
        "faq_database": 0.8       // Lower priority
    }
});

// Agent searches all collections
agent = aiAgent(
    name: "MultiSourceBot",
    memory: memory
);

Testing Your Custom Vector Memory

Unit Test Example

class extends="testbox.system.BaseSpec" {

    function run() {
        describe( "ElasticsearchVectorMemory", () => {

            beforeEach( () => {
                variables.memory = new ElasticsearchVectorMemory().configure({
                    collection: "test_memory",
                    esHost: "localhost",
                    esPort: 9200,
                    embeddingProvider: "openai",
                    embeddingModel: "text-embedding-3-small",
                    dimensions: 1536,
                    metric: "cosine"
                });
                variables.memory.key( "test-session-#createUUID()#" );
            });

            afterEach( () => {
                variables.memory.clear();
            });

            it( "can add and retrieve messages", () => {
                // Add messages
                memory.add( "BoxLang is a modern JVM language" );
                memory.add( "It has excellent Java interop" );
                memory.add( "The weather is sunny today" );

                expect( memory.count() ).toBe( 3 );

                // Search for relevant content
                var results = memory.getRelevant( "Tell me about BoxLang", 2 );

                expect( results ).toBeArray();
                expect( results.len() ).toBeGTE( 1 );
                expect( results[ 1 ].content ).toInclude( "BoxLang" );
            });

            it( "returns results sorted by relevance", () => {
                memory.add( "Python is a programming language" );
                memory.add( "BoxLang is built on the JVM" );
                memory.add( "Java runs on the JVM" );

                var results = memory.getRelevant( "JVM languages", 3 );

                // First result should be most relevant
                expect( results[ 1 ].score ).toBeGT( results[ 2 ].score );
            });

            it( "can clear all messages", () => {
                memory.add( "Test message 1" );
                memory.add( "Test message 2" );

                expect( memory.count() ).toBe( 2 );

                memory.clear();

                expect( memory.count() ).toBe( 0 );
            });

            it( "isolates messages by key", () => {
                memory.key( "session-1" );
                memory.add( "Message in session 1" );

                memory.key( "session-2" );
                memory.add( "Message in session 2" );

                memory.key( "session-1" );
                var session1Messages = memory.getAll();

                expect( session1Messages.len() ).toBe( 1 );
                expect( session1Messages[ 1 ].content ).toBe( "Message in session 1" );
            });

        });
    }
}

Best Practices

1. Always Call `super.configure()`

function configure( required struct config ) {
    // CRITICAL: Call parent configuration first
    super.configure( arguments.config );

    // Then add your custom configuration
    variables.myCustomSetting = arguments.config.myCustomSetting;

    return this;
}

2. Validate Configuration

function configure( required struct config ) {
    super.configure( arguments.config );

    // Validate required fields
    if ( !arguments.config.keyExists( "apiEndpoint" ) ) {
        throw(
            type: "ConfigurationException",
            message: "CustomVectorMemory requires 'apiEndpoint' in configuration"
        );
    }

    // Validate dimensions match embedding model
    if ( variables.dimensions != getModelDimensions( variables.embeddingModel ) ) {
        throw(
            type: "ConfigurationException",
            message: "Dimensions (#variables.dimensions#) don't match model (#variables.embeddingModel#)"
        );
    }

    return this;
}

3. Handle Errors Gracefully

function getRelevant( required string query, numeric limit = 5 ) {
    try {
        // Attempt vector search
        var embedding = generateEmbedding( arguments.query );
        return searchVectors( embedding, arguments.limit );

    } catch ( "EmbeddingException" e ) {
        // Log error and return empty results
        writeLog(
            type: "error",
            file: "vector-memory",
            text: "Failed to generate embedding: #e.message#"
        );
        return [];

    } catch ( "VectorSearchException" e ) {
        // Fallback to text search if vector search fails
        writeLog(
            type: "warning",
            file: "vector-memory",
            text: "Vector search failed, using fallback: #e.message#"
        );
        return fallbackTextSearch( arguments.query, arguments.limit );
    }
}

4. Optimize Embedding Generation

/**
 * Generate embeddings with batching and caching
 */
private function generateEmbeddings( required array texts ) {
    // Batch embed for efficiency
    if ( arguments.texts.len() > 1 ) {
        var result = aiEmbed(
            input: arguments.texts,  // Batch input
            provider: variables.embeddingProvider,
            model: variables.embeddingModel
        );
        return result.embeddings;
    } else {
        var result = aiEmbed(
            input: arguments.texts[ 1 ],
            provider: variables.embeddingProvider,
            model: variables.embeddingModel
        );
        return [ result.embeddings[ 1 ] ];
    }
}

5. Implement Proper Export/Import

function export() {
    return {
        "type": "custom",
        "class": getMetadata( this ).name,
        "key": variables.key,
        "config": {
            "collection": variables.collection,
            "embeddingProvider": variables.embeddingProvider,
            "embeddingModel": variables.embeddingModel,
            "dimensions": variables.dimensions,
            "metric": variables.metric
        },
        "messages": getAll(),
        "metadata": {
            "count": count(),
            "exportedAt": now()
        }
    };
}

function import( required struct data ) {
    // Validate import data
    if ( !arguments.data.keyExists( "messages" ) ) {
        throw( "Import data must contain 'messages' array" );
    }

    // Clear existing data
    clear();

    // Import messages
    arguments.data.messages.each( msg => {
        add( msg );
    } );

    return this;
}

6. Monitor Performance

function getRelevant( required string query, numeric limit = 5 ) {
    var startTime = getTickCount();

    try {
        var results = performVectorSearch( arguments.query, arguments.limit );

        var duration = getTickCount() - startTime;

        // Log slow queries
        if ( duration > 1000 ) {  // > 1 second
            writeLog(
                type: "warning",
                file: "vector-memory",
                text: "Slow vector search (#duration#ms) for query: #left(arguments.query, 50)#"
            );
        }

        return results;

    } catch ( any e ) {
        writeLog(
            type: "error",
            file: "vector-memory",
            text: "Vector search failed after #getTickCount() - startTime#ms: #e.message#"
        );
        rethrow;
    }
}

7. Support Metadata Filtering

function getRelevant(
    required string query,
    numeric limit = 5,
    struct filters = {}
) {
    var embedding = generateEmbedding( arguments.query );

    // Build query with metadata filters
    var searchQuery = {
        "vector": embedding,
        "limit": arguments.limit
    };

    // Add metadata filters
    if ( !arguments.filters.isEmpty() ) {
        searchQuery.filters = arguments.filters;
    }

    return executeSearch( searchQuery );
}

Common Patterns

Pattern 1: Wrapper Pattern

Wrap existing memory to add functionality:

class extends="BaseVectorMemory" {
    property name="wrappedMemory";

    function configure( required struct config ) {
        super.configure( arguments.config );
        variables.wrappedMemory = arguments.config.wrappedMemory;
        return this;
    }

    // Add functionality before/after delegation
    function add( required any message ) {
        preProcess( arguments.message );
        variables.wrappedMemory.add( arguments.message );
        postProcess( arguments.message );
        return this;
    }
}

Pattern 2: Adapter Pattern

Adapt existing clients to IVectorMemory interface:

class extends="BaseVectorMemory" {
    property name="thirdPartyClient";

    function configure( required struct config ) {
        super.configure( arguments.config );
        variables.thirdPartyClient = createThirdPartyClient( arguments.config );
        return this;
    }

    function add( required any message ) {
        // Adapt to third-party API
        variables.thirdPartyClient.insertDocument(
            collection: variables.collection,
            document: adaptMessage( arguments.message )
        );
        return this;
    }
}

Pattern 3: Composite Pattern

Combine multiple vector memories:

class extends="BaseVectorMemory" {
    property name="memories" type="array";

    function configure( required struct config ) {
        super.configure( arguments.config );
        variables.memories = arguments.config.memories ?: [];
        return this;
    }

    function add( required any message ) {
        variables.memories.each( memory => {
            memory.add( arguments.message );
        } );
        return this;
    }

    function getRelevant( required string query, numeric limit = 5 ) {
        var allResults = [];

        variables.memories.each( memory => {
            allResults.append( memory.getRelevant( arguments.query, arguments.limit ), true );
        } );

        return mergeAndRank( allResults, arguments.limit );
    }
}

Vector Memory Overview - Learn about built-in vector memory types
Custom Memory - Create custom standard memory implementations
Memory Systems - Understanding memory in BoxLang AI
Embeddings - Working with vector embeddings

Next Steps

Start Simple: Begin with BoxVector or extend an existing provider
Test Thoroughly: Write comprehensive unit tests
Monitor Performance: Track query times and cache hit rates
Optimize: Add caching, batching, and connection pooling
Document: Provide clear usage examples and configuration options

Need Help?

Community: BoxLang Discord
Documentation: BoxLang AI Docs
Examples: See /examples/vector-memory/ for working examples

PreviousBuilding Custom Document Loaders

Last updated 2 days ago

Good afternoon

hashtag🏗️ Custom Vector Memory Architecture

hashtag🎯 When to Build Custom Vector Memory

hashtag📚 Understanding BaseVectorMemory

hashtagWhat BaseVectorMemory Provides

hashtagWhat You Need to Implement

hashtagKey Properties in BaseVectorMemory

hashtag🔌 IVectorMemory Interface

hashtagExample 1: ElasticsearchVectorMemory

hashtagUsage Example

hashtagExample 2: RedisVectorMemory

hashtagExample 3: CachedVectorMemory

hashtagUsage Example

hashtagExample 4: MultiCollectionVectorMemory

hashtagUsage Example

hashtagTesting Your Custom Vector Memory

hashtagUnit Test Example

hashtagBest Practices

hashtag1. Always Call super.configure()

hashtag2. Validate Configuration

hashtag3. Handle Errors Gracefully

hashtag4. Optimize Embedding Generation

hashtag5. Implement Proper Export/Import

hashtag6. Monitor Performance

hashtag7. Support Metadata Filtering

hashtagCommon Patterns

hashtagPattern 1: Wrapper Pattern

hashtagPattern 2: Adapter Pattern

hashtagPattern 3: Composite Pattern

hashtagRelated Documentation

hashtagNext Steps

hashtagNeed Help?