Memory

Multi-tiered memory system with short-term, long-term, entity, and user-specific memory.

Quick Start

from praisonaiagents import Agent

# Enable memory with defaults
agent = Agent(instructions="You are a helpful assistant", memory=True)

# Use preset
agent = Agent(instructions="...", memory="redis")

# Use URL
agent = Agent(instructions="...", memory="redis://localhost:6379")

Usage Forms Table

Form	Example	When to Use
Bool	`memory=True`	Enable with defaults (SQLite)
String preset	`memory="redis"`	Use predefined backend
URL	`memory="redis://localhost:6379"`	Backend-specific connection
Dict	`memory={"backend": "postgres", "user_id": "u1"}`	Custom config
Array + overrides	`memory=["redis", {"user_id": "u1"}]`	Preset + customization
Config instance	`memory=MemoryConfig(backend="redis")`	Full control

Presets & Options

Preset	Backend	Description
`"file"`	Local file	File-based storage
`"sqlite"`	SQLite	Default, local database
`"redis"`	Redis	Fast, distributed
`"postgres"`	PostgreSQL	Scalable, persistent
`"postgresql"`	PostgreSQL	Alias for postgres
`"mongodb"`	MongoDB	Document-based
`"mem0"`	Mem0	External memory service

URL Schemes

Scheme	Backend
`postgresql://`	PostgreSQL
`postgres://`	PostgreSQL
`redis://`	Redis
`rediss://`	Redis with SSL
`sqlite://`	SQLite
`mongodb://`	MongoDB
`mongodb+srv://`	MongoDB Atlas

Precedence Ladder

Resolution Order: Instance > Config > Array > Dict > String > Bool > DefaultWhen you pass memory=, the resolver checks in this order:

Instance - MemoryManager instance? Use as-is
Config - MemoryConfig instance? Use as-is
Array - ["preset", {"override": value}]? Apply overrides
Dict - {"key": value}? Convert to config
String - URL or preset? Parse URL or look up preset
Bool - True? Use defaults. False? Disable

Memory Tiers

Tier	Purpose	Storage	Retention
Short-Term (STM)	Active conversation	SQLite	Ephemeral
Long-Term (LTM)	Persistent knowledge	SQLite + Vector	Permanent
Entity Memory	Structured entities	LTM subset	Permanent
User Memory	User preferences	LTM with user_id	Permanent

Classes

MemoryClient

Method	Description
`store_in_short_memory(memory, ...)`	Store in STM
`store_in_long_memory(memory, ...)`	Store in LTM
`search_memories(query, ...)`	Search across tiers
`get_short_memories(...)`	Retrieve STM
`get_long_memories(...)`	Retrieve LTM
`clear_short_memory(...)`	Clear STM
`clear_long_memory(...)`	Clear LTM

Quality Scoring Functions

compute_quality_score

Calculate overall quality score from individual metrics.

def compute_quality_score(
    completeness: float,
    relevance: float,
    clarity: float,
    accuracy: float,
    weights: Optional[Dict[str, float]] = None
) -> float

calculate_quality_metrics

Use LLM to evaluate output quality against expectations.

def calculate_quality_metrics(
    output: str,
    expected_output: str,
    llm=None,
    custom_prompt: Optional[str] = None
) -> Dict[str, float]

Memory Tiers

1. Short-Term Memory (STM)

Purpose: Immediate context and active conversation
Storage: SQLite database (.praison/short_term.db)
Retention: Ephemeral, cleared between sessions
Use Cases: Current task context, recent interactions

2. Long-Term Memory (LTM)

Purpose: Persistent knowledge across sessions
Storage: SQLite + optional vector store
Retention: Permanent with quality filtering
Use Cases: Learned facts, important outcomes

3. Entity Memory

Purpose: Structured information about entities
Storage: Subset of LTM with special formatting
Format: Entity {name}({type}): {desc} | relationships: {relations}
Use Cases: People, organizations, locations

4. User Memory

Purpose: User-specific preferences and history
Storage: LTM with user_id filtering
Isolation: Strict user separation
Use Cases: Personalization, preferences

Configuration

Basic Configuration

config = {
    "provider": "rag",  # Options: rag, mem0, none
    "use_embedding": True,
    "rag_db_path": ".praison/chroma"
}

Advanced Configuration with Graph Support

config = {
    "provider": "mem0",
    "config": {
        "graph_store": {
            "provider": "neo4j",
            "config": {
                "url": "neo4j+s://your-instance.neo4j.io",
                "username": "neo4j",
                "password": "your-password"
            }
        },
        "vector_store": {
            "provider": "chroma",
            "config": {
                "collection_name": "agent_memory"
            }
        },
        "llm": {
            "provider": "openai",
            "config": {
                "model": "gpt-4o-mini"
            }
        }
    }
}

Graph Database Options

Neo4j Configuration

"graph_store": {
    "provider": "neo4j",
    "config": {
        "url": "neo4j+s://...",
        "username": "neo4j",
        "password": "..."
    }
}

Memgraph Configuration

"graph_store": {
    "provider": "memgraph",
    "config": {
        "url": "bolt://localhost:7687",
        "username": "memgraph",
        "password": "..."
    }
=======
title: "Memory"
sidebarTitle: "Memory"
description: "Multi-tiered memory system with graph support for intelligent agents"
icon: "brain"
---

## Overview

The Memory module provides a sophisticated multi-tiered memory system that enables agents to maintain context across conversations, store and retrieve information efficiently, and even utilize graph databases for complex relationship mapping.

```mermaid
flowchart TB
    subgraph "Memory Tiers"
        STM[Short-term Memory<br/>Recent Interactions]
        LTM[Long-term Memory<br/>Important Information]
        ETM[Entity Memory<br/>People, Places, Things]
        USM[User Memory<br/>Preferences & History]
    end
    
    subgraph "Storage Backends"
        SQL[SQLite<br/>Local Storage]
        VEC[ChromaDB<br/>Vector Storage]
        MEM[Mem0<br/>Managed Service]
        GRA[Neo4j/Memgraph<br/>Graph Database]
    end
    
    subgraph "Quality Management"
        QS[Quality Scoring<br/>0.0 - 1.0]
        QF[Quality Filtering<br/>Threshold: 0.7]
    end
    
    STM --> SQL
    LTM --> VEC
    ETM --> GRA
    USM --> SQL
    
    STM --> QS
    LTM --> QS
    QS --> QF
    
    style STM fill:#189AB4,color:#fff
    style LTM fill:#2E8B57,color:#fff
    style GRA fill:#8B0000,color:#fff
    style QS fill:#DAA520,color:#fff

Quick Start

Install with memory support

pip install praisonaiagents

Basic memory usage

from praisonaiagents import Memory

# Initialize memory system
memory = Memory()

# Store information
memory.add(
    text="The user prefers Python over JavaScript",
    memory_type="long"
)

# Search memories
results = memory.search("programming preferences")

# Build context for agents
context = memory.build_context_for_task(
    task_description="Help with a Python project",
    max_items=5
)

Advanced graph memory

from praisonaiagents import Memory

# Initialize with graph support
memory = Memory(
    graph_enabled=True,
    graph_uri="bolt://localhost:7687",
    graph_user="neo4j",
    graph_password="password"
)

# Store entity relationships
memory.add(
    text="John works at TechCorp as a Senior Developer",
    memory_type="entity"
)

# Query graph relationships
results = memory.search("Who works at TechCorp?")

Configuration

Basic Configuration

memory = Memory(
    provider="rag",  # Options: "rag", "mem0", "none"
    use_embedding=True,
    api_key=None,  # For mem0 provider
    memory={"user_id": "default_user"},
    debug=False
)

Advanced Configuration

memory = Memory(
    # Storage paths
    rag_db_path="custom/path/chroma_db",
    short_db="custom/path/short_term.db",
    long_db="custom/path/long_term.db",
    entity_db="custom/path/entity.db",
    user_db="custom/path/user.db",
    
    # Graph configuration
    graph_enabled=True,
    graph_uri="bolt://localhost:7687",
    graph_user="neo4j",
    graph_password="password",
    
    # Quality settings
    quality_threshold=0.7,
    
    # Debug mode
    debug=True
)

Memory Types

API Reference

Constructor

Memory(
    provider: str = "rag",
    use_embedding: bool = True,
    api_key: Optional[str] = None,
    user_id: str = "default_user",
    rag_db_path: str = "memory_chroma_db",
    short_db: str = "short_term_memory.db",
    long_db: str = "long_term_memory.db",
    entity_db: str = "entity_memory.db",
    user_db: str = "user_memory.db",
    graph_enabled: bool = False,
    graph_uri: Optional[str] = None,
    graph_user: Optional[str] = None,
    graph_password: Optional[str] = None,
    quality_threshold: float = 0.7,
    debug: bool = False
)

Core Methods

add()

Store information in memory with optional quality scoring.

add(
    text: str,
    memory_type: str = "short",
    quality_score: Optional[float] = None,
    metadata: Optional[Dict[str, Any]] = None
) -> None

Parameters:

text - Content to store
memory_type - Type: “short”, “long”, “entity”, or “user”
quality_score - Quality rating (0.0-1.0, auto-calculated if None)
metadata - Additional metadata

search()

Search across all memory types for relevant information.

search(
    query: str,
    memory_type: Optional[str] = None,
    limit: int = 5
) -> List[Dict[str, Any]]

Returns list of:

{
    'text': str,          # Memory content
    'memory_type': str,   # Type of memory
    'timestamp': str,     # Creation time
    'quality_score': float,  # Quality rating
    'metadata': dict      # Additional data
}

update()

Update an existing memory entry.

update(
    memory_id: str,
    text: str,
    memory_type: str = "short",
    quality_score: Optional[float] = None
) -> None

delete()

Delete a specific memory entry.

delete(
    memory_id: str,
    memory_type: str = "short"
) -> None

Context Building

build_context_for_task()

Build formatted context for a specific task.

build_context_for_task(
    task_description: str,
    max_items: int = 10
) -> str

Example output:

Based on memory:
- User prefers Python for data science projects
- Previous experience with pandas and numpy
- Interested in machine learning applications

get_context()

Get all memories formatted as context.

get_context(
    memory_type: Optional[str] = None,
    limit: int = 10
) -> str

Quality Management

calculate_quality_score()

Calculate quality score for a memory entry.

calculate_quality_score(
    text: str,
    memory_type: str = "short"
) -> float

Scoring factors:

Information density
Specificity
Relevance indicators
Entity mentions
Temporal relevance

get_quality_memories()

Retrieve only high-quality memories.

get_quality_memories(
    memory_type: str = "long",
    min_quality: float = 0.7,
    limit: int = 10
) -> List[Dict[str, Any]]

Utility Methods

get_memories()

Retrieve raw memories from storage.

get_memories(
    memory_type: str = "short",
    limit: int = 10
) -> List[Dict[str, Any]]

clear()

Clear all memories of a specific type.

clear(memory_type: str = "all") -> None

Options: “short”, “long”, “entity”, “user”, “all”

get_stats()

Get memory system statistics.

get_stats() -> Dict[str, Any]

Returns:

{
    'total_memories': int,
    'by_type': {
        'short': int,
        'long': int,
        'entity': int,
        'user': int
    },
    'quality_distribution': dict,
    'storage_info': dict

}

Usage Examples

Basic Memory Operations

from praisonaiagents import MemoryClient

# Initialize memory
memory = MemoryClient(verbose=2)

# Store memories
memory.store_in_short_memory("User asked about pricing")
memory.store_in_long_memory(
    "Company pricing: $99/month for pro plan",
    quality_score=0.9
)

# Search memories
results = memory.search_memories("pricing", k=5)

Quality-Based Storage

from praisonaiagents import calculate_quality_metrics, compute_quality_score

# Evaluate output quality
metrics = calculate_quality_metrics(
    output="Generated report content...",
    expected_output="Comprehensive analysis with data"
)

# Calculate overall score
score = compute_quality_score(
    completeness=metrics['completeness'],
    relevance=metrics['relevance'],
    clarity=metrics['clarity'],
    accuracy=metrics['accuracy']
)

# Store only high-quality outputs
if score > 0.7:
    memory.store_in_long_memory(output, quality_score=score)

Entity Memory Management

# Store entity information
memory.store_entity_memory(
    name="Acme Corp",
    entity_type="Company",
    desc="Leading provider of AI solutions",
    relations=["Founded by John Doe", "Partnered with TechCo"],
    memory={"user_id": "user123"}
)

# Retrieve entity context
context = memory.build_context_for_task(
    "Tell me about Acme Corp's partnerships"
)

Agent Integration

from praisonaiagents import Agent

# Agent with memory
agent = Agent(
    name="Assistant",
    role="Helpful AI assistant",
    memory={
        "provider": "rag",
        "use_embedding": True
    }
)

# Memory is automatically managed during conversations
response = agent.chat("Remember that my favorite color is blue")

Graph-Enhanced Memory

# Configure with Neo4j
memory = MemoryClient({
    "provider": "mem0",
    "config": {
        "graph_store": {
            "provider": "neo4j",
            "config": {...}
        }
    }
})

# Store with relationships
memory.memory.add(
    "John Doe is the CEO of Acme Corp",
    metadata={"entities": ["John Doe", "Acme Corp"]}
)

# Graph-aware search
results = memory.search_memories(
    "Who works at Acme Corp?",
    rerank=True
)

Quality Metrics

Completeness (0-1)

How thoroughly the content addresses the requirements.

Relevance (0-1)

How well the content matches the expected output.

Clarity (0-1)

How clear and well-structured the content is.

Accuracy (0-1)

Factual correctness of the information.

Custom Weights

weights = {
    "completeness": 0.3,
    "relevance": 0.4,
    "clarity": 0.2,
    "accuracy": 0.1
}
score = compute_quality_score(**metrics, weights=weights)

Best Practices

Use Quality Filtering - Set appropriate min_quality thresholds
Scope Memories - Use user_id and agent_id for proper isolation
Regular Cleanup - Clear short-term memory between sessions
Graph for Relationships - Use graph stores for complex entity relationships
Monitor Storage - Check database sizes periodically
Test Retrieval - Verify context building produces relevant results =======

Basic Memory Usage

from praisonaiagents import Agent, Memory

# Create memory system
memory = Memory(memory={"user_id": "user_123"})

# Create agent with memory
agent = Agent(
    name="Assistant",
    role="Personal AI Assistant",
    memory=memory
)

# Conversation that builds memory
agent.chat("I prefer morning meetings")
# Automatically stored in memory

agent.chat("Schedule something for tomorrow")
# Uses memory: "Scheduling morning meeting as you prefer"

Quality-Based Storage

from praisonaiagents import Memory

memory = Memory(quality_threshold=0.8)

# High-quality information
memory.add(
    "User's API key: sk-abc123def456",
    memory_type="long",
    quality_score=0.9
)

# Low-quality information (won't be stored in long-term)
memory.add(
    "User said hello",
    memory_type="long",
    quality_score=0.3
)

# Get only high-quality memories
important = memory.get_quality_memories(min_quality=0.8)

Graph Memory Example

from praisonaiagents import Memory

# Setup graph memory
memory = Memory(
    graph_enabled=True,
    graph_uri="bolt://localhost:7687",
    graph_user="neo4j",
    graph_password="password"
)

# Store entity relationships
memory.add(
    "Alice manages Bob and Charlie at DataCorp",
    memory_type="entity"
)

memory.add(
    "DataCorp acquired SmallStartup in 2024",
    memory_type="entity"
)

# Query relationships
results = memory.search("Who does Alice manage?")
# Returns information about Bob and Charlie

results = memory.search("DataCorp acquisitions")
# Returns information about SmallStartup acquisition

from praisonaiagents import Memory, Agent, AgentTeam

# Shared memory system
shared_memory = Memory(config={"provider": "rag"})

# Create agents with shared memory
researcher = Agent(
    name="Researcher",
    role="Research Analyst",
    memory=shared_memory
)

writer = Agent(
    name="Writer",
    role="Content Creator",
    memory=shared_memory
)

# Researcher stores findings
researcher.chat("Found that 73% of users prefer dark mode")

# Writer can access the same memory
response = writer.chat("Write about user preferences")
# Uses the 73% statistic from shared memory

Best Practices

Memory Type Selection

Short-term: Conversation context, temporary state Long-term: Facts, preferences, important information Entity: People, places, organizations, relationships User: Personal data, settings, history

Quality Management

Set appropriate quality thresholds (0.7-0.8 recommended)
Manually score critical information higher
Periodically review and clean low-quality memories
Use quality scores for retrieval filtering

Performance Optimization

Limit memory searches to necessary types
Use appropriate search limits
Clear short-term memory periodically
Index frequently accessed memories

Privacy & Security

Separate user memories by user_id
Avoid storing sensitive data in plain text
Implement access controls for shared memory
Regular cleanup of old user data

Provider Comparison

Troubleshooting

Common Issues:

Advanced Configuration

Custom Quality Scoring

from praisonaiagents import Memory

class CustomMemory(Memory):
    def calculate_quality_score(self, text: str, memory_type: str) -> float:
        # Base score
        score = super().calculate_quality_score(text, memory_type)
        
        # Custom adjustments
        if "important" in text.lower():
            score += 0.2
        if len(text) > 200:  # Favor detailed information
            score += 0.1
        if memory_type == "entity":  # Boost entity memories
            score += 0.15
            
        return min(score, 1.0)

# Use custom memory
memory = CustomMemory(quality_threshold=0.75)

Memory Middleware

from praisonaiagents import Agent, Memory
from datetime import datetime

class AuditedMemory(Memory):
    def add(self, text: str, memory_type: str = "short", **kwargs):
        # Add audit metadata
        kwargs['metadata'] = kwargs.get('metadata', {})
        kwargs['metadata']['added_at'] = datetime.now().isoformat()
        kwargs['metadata']['source'] = 'agent_conversation'
        
        # Log addition
        print(f"[AUDIT] Adding {memory_type} memory: {text[:50]}...")
        
        super().add(text, memory_type, **kwargs)

# Use with agents
memory = AuditedMemory()
agent = Agent(name="Audited", memory=memory)

Summary

The Memory module provides a comprehensive solution for agent memory management: ✅ Multi-tiered Architecture - Different memory types for different needs
✅ Quality Management - Automatic scoring and filtering
✅ Graph Support - Complex relationship mapping with Neo4j/Memgraph
✅ Flexible Storage - Multiple backend options
✅ Context Building - Automatic context generation for tasks Perfect for building agents that:

Maintain conversation context
Remember user preferences
Track entity relationships
Build knowledge over time

Guide

Reference

​Memory

​Quick Start

​Usage Forms Table

​Presets & Options

​URL Schemes

​Precedence Ladder

​Memory Tiers

​Classes

​MemoryClient

​Quality Scoring Functions

​compute_quality_score

​calculate_quality_metrics

​Memory Tiers

​1. Short-Term Memory (STM)

​2. Long-Term Memory (LTM)

​3. Entity Memory

​4. User Memory

​Configuration

​Basic Configuration

​Advanced Configuration with Graph Support

​Graph Database Options

​Neo4j Configuration

​Memgraph Configuration

​Quick Start

​Configuration

​Basic Configuration

​Advanced Configuration

​Memory Types

​API Reference

​Constructor

​Core Methods

​add()

​search()

​update()

​delete()

​Context Building

​build_context_for_task()

​get_context()

​Quality Management

​calculate_quality_score()

​get_quality_memories()

​Utility Methods

​get_memories()

​clear()

​get_stats()

​Usage Examples

​Basic Memory Operations

​Quality-Based Storage

​Entity Memory Management

​Agent Integration

​Graph-Enhanced Memory

​Quality Metrics

​Completeness (0-1)

​Relevance (0-1)

​Clarity (0-1)

​Accuracy (0-1)

​Custom Weights

​Best Practices

​Basic Memory Usage

​Quality-Based Storage

​Graph Memory Example

​Multi-Agent Memory Sharing

​Best Practices

Memory Type Selection

Quality Management

Performance Optimization

Privacy & Security

​Provider Comparison

​Troubleshooting

​Advanced Configuration

​Custom Quality Scoring

​Memory Middleware

​Summary

Memory

Quick Start

Usage Forms Table

Presets & Options

URL Schemes

Precedence Ladder

Memory Tiers

Classes

MemoryClient

Quality Scoring Functions

compute_quality_score

calculate_quality_metrics

Memory Tiers

1. Short-Term Memory (STM)

2. Long-Term Memory (LTM)

3. Entity Memory

4. User Memory

Configuration

Basic Configuration

Advanced Configuration with Graph Support

Graph Database Options

Neo4j Configuration

Memgraph Configuration

Quick Start

Configuration

Basic Configuration

Advanced Configuration

Memory Types

API Reference

Constructor

Core Methods

add()

search()

update()

delete()

Context Building

build_context_for_task()

get_context()

Quality Management

calculate_quality_score()

get_quality_memories()

Utility Methods

get_memories()

clear()

get_stats()

Usage Examples

Basic Memory Operations

Quality-Based Storage

Entity Memory Management

Agent Integration

Graph-Enhanced Memory

Quality Metrics

Completeness (0-1)

Relevance (0-1)

Clarity (0-1)

Accuracy (0-1)

Custom Weights

Best Practices

Basic Memory Usage

Quality-Based Storage

Graph Memory Example

Multi-Agent Memory Sharing

Best Practices

Provider Comparison

Troubleshooting

Advanced Configuration

Custom Quality Scoring

Memory Middleware

Summary