📊Agent Memory Presentation | 🎥AIEWF Richmond's Talk

Overview

Key Features

Installation

pip install memorizz

Prerequisites

Quick Start

1. Basic MemAgent Setup

importosfrommemorizz.memory_provider.mongodb.providerimportMongoDBConfig,MongoDBProviderfrommemorizz.memagentimportMemAgentfrommemorizz.llms.openaiimportOpenAI# Set up your API keysos.environ["OPENAI_API_KEY"]="your-openai-api-key"# Configure MongoDB memory providermongodb_config=MongoDBConfig(uri="your-mongodb-atlas-uri")memory_provider=MongoDBProvider(mongodb_config)# Create a MemAgentagent=MemAgent(model=OpenAI(model="gpt-4"),instruction="You are a helpful assistant with persistent memory.",memory_provider=memory_provider)# Start conversing - the agent will remember across sessionsresponse=agent.run("Hello! My name is John and I'm a software engineer.")print(response)# Later in another session...response=agent.run("What did I tell you about myself?")print(response)# Agent remembers John is a software engineer

Table of single agent and multi-agent setups, their descriptions, and links to example notebooks

Memory System Components and Examples

2. Creating Specialized Agents with Personas

frommemorizz.long_term_memory.semantic.personaimportPersonafrommemorizz.long_term_memory.semantic.persona.role_typeimportRoleType# Create a technical expert persona using predefined role typestech_expert=Persona(name="TechExpert",role=RoleType.TECHNICAL_EXPERT,# Use predefined role enumgoals="Help developers solve complex technical problems with detailed explanations.",background="10+ years experience in Python, AI/ML, and distributed systems.")# Apply persona to agentagent.set_persona(tech_expert)agent.save()# Now the agent will respond as a technical expertresponse=agent.run("How should I design a scalable microservices architecture?")

3. Tool Registration and Function Calling

frommemorizz.databaseimportMongoDBTools,MongoDBToolsConfigfrommemorizz.embeddings.openaiimportget_embedding# Configure tools databasetools_config=MongoDBToolsConfig(mongo_uri="your-mongodb-atlas-uri",db_name="my_tools_db",get_embedding=get_embedding# Required embedding function)# Register tools using decoratorwithMongoDBTools(tools_config)astools:toolbox=tools.mongodb_toolbox()@toolboxdefcalculate_compound_interest(principal:float,rate:float,time:int)->float:"""Calculate compound interest for financial planning."""returnprincipal* (1+rate)**time@toolboxdefget_weather(city:str)->str:"""Get current weather for a city."""# Your weather API integration herereturnf"Weather in{city}: 72°F, sunny"# Add tools to your agentagent.add_tool(toolbox=toolbox)# Agent can now discover and use these tools automaticallyresponse=agent.run("What's the weather in San Francisco and calculate interest on $1000 at 5% for 3 years?")

4. Semantic Cache for Performance Optimization

# Enable semantic cache on any MemAgentagent=MemAgent(model=OpenAI(model="gpt-4"),instruction="You are a helpful assistant.",memory_provider=memory_provider,semantic_cache=True,# Enable semantic cachesemantic_cache_config={"similarity_threshold":0.85,# Adjust sensitivity (0.0-1.0)"max_cache_size":1000,# Maximum cached responses"ttl_hours":24.0# Cache expiration time    })# Similar queries will use cached responsesresponse1=agent.run("What is the capital of France?")response2=agent.run("Tell me France's capital city")# Cache hit! ⚡response3=agent.run("What's the capital of Japan?")# New query, cache miss# For external frameworks - use standalone semantic cachefrommemorizz.short_term_memory.semantic_cacheimportStandaloneSemanticCachecache=StandaloneSemanticCache(similarity_threshold=0.8,embedding_provider="openai")# Integrate with any agent frameworkdefmy_agent(query:str)->str:# Check cache firstcached_response=cache.query(query)ifcached_response:returncached_response# Generate new response with your LLMresponse=your_llm_call(query)# Cache for future similar queriescache.cache_response(query,response)returnresponse# Memory-scoped caching for multi-session isolationfrommemorizz.short_term_memory.semantic_cacheimportcreate_semantic_cacheuser_cache=create_semantic_cache(agent_id="assistant",memory_id="user_session_123",# Isolate cache per user sessionsimilarity_threshold=0.9)

Core Concepts

Memory Types

Long-Term Knowledge Management

# Add knowledge to long-term memoryknowledge_id=agent.add_long_term_memory("I prefer Python for backend development due to its simplicity and extensive libraries.",namespace="preferences")# Retrieve related knowledgeknowledge_entries=agent.retrieve_long_term_memory(knowledge_id)# Update existing knowledgeagent.update_long_term_memory(knowledge_id,"I prefer Python for backend development and FastAPI for building APIs.")# Delete knowledge when no longer neededagent.delete_long_term_memory(knowledge_id)

Tool Discovery

# Tools are automatically found based on intentagent.run("I need to check the weather")# Finds and uses get_weather toolagent.run("Help me calculate some financial returns")# Finds compound_interest tool

Advanced Usage

Custom Memory Providers

frommemorizz.memory_provider.baseimportMemoryProviderclassCustomMemoryProvider(MemoryProvider):defstore(self,data,memory_store_type):# Your custom storage logicpassdefretrieve_by_query(self,query,memory_store_type,limit=10):# Your custom retrieval logicpass

Multi-Agent Workflows

# Create specialized delegate agentsdata_analyst=MemAgent(model=OpenAI(model="gpt-4"),instruction="You are a data analysis expert.",memory_provider=memory_provider)report_writer=MemAgent(model=OpenAI(model="gpt-4"),instruction="You are a report writing specialist.",memory_provider=memory_provider)# Create orchestrator agent with delegatesorchestrator=MemAgent(model=OpenAI(model="gpt-4"),instruction="You coordinate between specialists to complete complex tasks.",memory_provider=memory_provider,delegates=[data_analyst,report_writer])# Execute multi-agent workflowresponse=orchestrator.run("Analyze our sales data and create a quarterly report.")

Memory Management Operations

# Save agent state to memory provideragent.save()# Load existing agent by IDexisting_agent=MemAgent.load(agent_id="your-agent-id",memory_provider=memory_provider)# Update agent configurationagent.update(instruction="Updated instruction for the agent",max_steps=30)# Delete agent and optionally cascade delete memoriesMemAgent.delete_by_id(agent_id="agent-id-to-delete",cascade=True,# Deletes associated memoriesmemory_provider=memory_provider)

Architecture

┌─────────────────┐│   MemAgent      │  ← High-level agent interface├─────────────────┤│   Persona       │  ← Agent personality & behavior├─────────────────┤│   Toolbox       │  ← Function registration & discovery├─────────────────┤│ Memory Provider │  ← Storage abstraction layer├─────────────────┤│ Vector Search   │  ← Semantic similarity & retrieval├─────────────────┤│   MongoDB       │  ← Persistent storage backend└─────────────────┘

Examples

Configuration

MongoDB Atlas Setup

Environment Variables

# Requiredexport OPENAI_API_KEY="your-openai-api-key"export MONGODB_URI="your-mongodb-atlas-uri"# Optionalexport MONGODB_DB_NAME="memorizz"# Default database name

Troubleshooting

Contributing

License

Educational Resources