Advanced Memory System

Multi-tiered memory with quality scoring and graph support

Advanced Memory System

The advanced memory system provides sophisticated memory management with short-term, long-term, entity, and user-specific storage, enhanced by quality scoring and optional graph database support.

Key Features

Separate short-term and long-term memory systems

4-metric quality assessment for stored memories

User, agent, and run-specific memory scoping

Automatic entity extraction and storage

Optional Neo4j/Memgraph for relationships

Quality-based filtering and relevance ranking

Quick Start

Agent with Memory

from praisonaiagents import Agent, AgentTeam

# Single agent with memory (simplest)
agent = Agent(
    name="Assistant",
    instructions="You are a helpful assistant with memory.",
    memory=True  # Enable with defaults
)

# Agent with memory config
agent = Agent(
    name="Assistant",
    instructions="You are a helpful assistant with memory.",
    memory={
        "backend": "rag",
        "user_id": "user123",
        "use_embedding": True
    }
)

# Multi-agent with shared memory
agents = AgentTeam(
    agents=[agent1, agent2],
    memory={
        "backend": "rag",
        "use_embedding": True
    }
)

Memory Tiers

Short-term Memory (STM)

Short-term memory is cleared between sessions and used for immediate context.

# Store temporary context and capture the returned id
mid = memory.store_short_term(
    text="User is asking about Python programming",
    metadata={"user_id": "user123", "context": "current_conversation"}
)

# Search recent context
context = memory.search_short_term(
    query="Python",
    limit=5
)

# Targeted delete uses the id (returns True on success)
memory.delete_short_term(mid)

# Or wipe everything
memory.reset_short_term()

memory.reset_short_term() / .reset_long_term() now work on the default rag / chroma provider, not just sqlite / in_memory and dakera. On sqlite / in_memory providers, reset_* / delete_* delegate to the active adapter (mirroring search_short_term), so they operate on the adapter-created short_term_memory / long_term_memory tables.

On the default rag provider, resetting short-term also clears long-term — Chroma stores both tiers in one collection. Use DakeraMemoryAdapter if you need tier isolation.

Adapters that don’t implement reset_short_term / reset_long_term now log a warning stating the tier was NOT cleared, instead of silently succeeding — check logs after upgrade.

Since PraisonAI PR #4020, Chroma-backed memory (the default) supports reset. On earlier releases, Memory().reset_short_term() / .reset_long_term() was a documented, callable, exception-free API that quietly did nothing.

Long-term Memory (LTM)

Long-term memory persists across sessions and stores important information.

# Store persistent knowledge
memory.store_long_term(
    text="User works as a data scientist at TechCorp",
    memory={"user_id": "user123"},
    quality=0.95,
    metadata={"type": "user_info", "category": "professional"}
)

# Retrieve with quality filter — user_id is a first-class kwarg
memories = memory.search_long_term(
    query="user profession",
    user_id="user123",
    min_quality=0.8
)

Entity Memory

Entity memory stores information about specific people, places, or things.

# Store entity information
memory.store_entity(
    name="TechCorp",
    text="TechCorp is a leading AI company founded in 2020",
    entity_type="organization",
    metadata={"industry": "technology", "size": "large"}
)

# Search entity information
entity_info = memory.search_entity(
    name="TechCorp",
    query="company details"
)

User Memory

User memory stores personalised information and preferences.

# Store user preferences
memory.store_user_memory(
    memory={"user_id": "user123"},
    text="Prefers concise explanations with code examples",
    extra={"preference_type": "communication_style"}
)

# Retrieve user context
user_context = memory.search_user_memory(
    memory={"user_id": "user123"},
    query="preferences",
    limit=3
)

Configuration Options

RAG Configuration (Default)

memory_config = {
    "provider": "rag",
    "use_embedding": True,
    "short_db": ".praison/short_term.db",
    "long_db": ".praison/long_term.db",
    "entity_db": ".praison/entity.db",
    "rag_db_path": ".praison/chroma_db",
    "embedding_model": "text-embedding-3-small"
}

Mem0 Configuration

memory_config = {
    "provider": "mem0",
    "config": {
        "api_key": "your-mem0-api-key",
        "org_id": "your-org-id",
        "project_id": "your-project-id"
    }
}

Dakera Configuration

Requires: pip install "praisonaiagents[dakera]"

memory_config = {
    "provider": "dakera",
    "config": {
        "url": "http://localhost:3000",      # or DAKERA_URL / DAKERA_API_URL env var
        "api_key": "dk-...",                 # or DAKERA_API_KEY env var
        "agent_id": "my-agent",             # or DAKERA_AGENT_ID env var
        "short_term_type": "working",
        "long_term_type": "episodic",
        "default_importance": 0.5,
    }
}

See Dakera Memory for the full guide.

Graph Memory Configuration

memory_config = {
    "provider": "mem0",
    "config": {
        "graph_store": {
            "provider": "neo4j",
            "config": {
                "url": "bolt://localhost:7687",
                "username": "neo4j",
                "password": "password"
            }
        },
        "vector_store": {
            "provider": "qdrant",
            "config": {
                "host": "localhost",
                "port": 6333
            }
        },
        "version": "v1.1"
    }
}

Quality Scoring System

Quality Metrics

Completeness

Measures how complete and comprehensive the information is

Relevance

Measures how relevant the information is to the context

Clarity

Measures how clear and understandable the information is

Accuracy

Measures the factual accuracy of the information

Quality Calculation

# Default weights
weights = {
    "completeness": 0.25,
    "relevance": 0.35,
    "clarity": 0.20,
    "accuracy": 0.20
}

# Overall quality = weighted average
quality = (
    completeness * weights["completeness"] +
    relevance * weights["relevance"] +
    clarity * weights["clarity"] +
    accuracy * weights["accuracy"]
)

Local LLM for quality scoring

Memory quality scoring uses litellm when available and falls back to the OpenAI client. Point it at any OpenAI-compatible local server by either:

Via memory config (recommended when the rest of your memory layer is config-driven):

from praisonaiagents import Memory

memory = Memory(config={
    "llm": {
        "config": {
            "base_url": "http://localhost:11434/v1",
            "model": "ollama/qwen3:0.6b",
        }
    }
})

Via environment (works with any call site, not just memory):

export OPENAI_BASE_URL=http://localhost:11434/v1
export PRAISONAI_AUXILIARY_MODEL=ollama/qwen3:0.6b

Before PR #4812, the OpenAI-client fallback in calculate_quality_metrics silently ignored base_url supplied through config — it worked only if you also exported OPENAI_BASE_URL. If you were seeing memory reads/writes go to your local server but quality scoring still hit api.openai.com, that was the bug; it is now fixed.

Advanced Features

Scoped search with user_id and metadata_filter

search_short_term and search_long_term accept user_id and metadata_filter as first-class kwargs to scope results per tenant.

from praisonaiagents import Agent, Memory

memory = Memory(config={"provider": "rag", "use_embedding": True})
agent = Agent(
    name="TenantAssistant",
    instructions="Answer scoped to the current tenant.",
    memory=True,
)

# Multi-tenant memory scoping
results = memory.search_long_term(
    query="project details",
    user_id="tenant_a",             # tenant isolation — always wins
    metadata_filter={"project": "atlas"},  # extra scoping
    limit=10,
)

user_id is now enforced twice on every search — once at the Memory layer (existing) and again inside SearchMixin.search_short_term / SearchMixin.search_long_term as of PraisonAI PR #4819. The lower layer auto-derives the same user_id metadata filter, so adapters that accept user_id but ignore it in their own query (Chroma/RAG, SQLite, in-memory) still isolate per-tenant results. If a caller passes both an explicit user_id and a conflicting metadata_filter["user_id"], the explicit user_id always wins — a caller cannot widen scope to another tenant’s data.

user_id takes precedence over any metadata_filter["user_id"], so a route can never widen scope to another tenant. Applied across every backend — Mem0, MongoDB (with and without vector search), ChromaDB, memory-adapter, and the SQLite fallback. When a metadata_filter is present, each backend over-fetches by 10× so results ranked beyond limit still survive the post-filter, then truncates back to limit.

Direct SearchMixin use (subclassing or embedding in a custom store) previously depended on the adapter to filter by user_id. If you have such code, no change is required — the filter is applied automatically now — but you may safely remove any duplicate post-filtering you added as a workaround.

Before PR #3855, user_id and metadata_filter were silently absorbed into **kwargs and never applied — per-user scoping was broken on the class-body path. Pass them explicitly, not inside a memory={...} dict.

Context Building

Build comprehensive context for tasks:

# Automatically merge relevant memories
context = memory.build_context_for_task(
    task_descr="Write a Python tutorial",
    memory={"user_id": "user123"},
    additional="Focus on data science applications",
    max_items=5
)

# Context includes:
# - Relevant long-term memories
# - User preferences
# - Recent short-term context

Task Output Finalisation

Store task results with quality assessment:

from praisonaiagents.agent.task import TaskOutput

# Create task output
output = TaskOutput(
    raw="Tutorial content...",
    agent_name="Writer",
    task_description="Write Python tutorial"
)

# Finalise with quality scoring
memory.finalize_task_output(
    task_output=output,
    memory={"user_id": "user123"}
)

Memory Citations

Automatically cite memory sources:

# Memories include source information
result = memory.search_long_term("Python", user_id="user123")
# Returns: {
#     'text': 'Python is...',
#     'metadata': {'source': 'conversation_2024_01_15', ...}
# }

Memory Reranking

Enhance search results with intelligent reranking based on relevance scores:

from praisonaiagents import Agent, AgentTeam

# Create agent
agent = Agent(
    name="Researcher",
    role="Research assistant"
)

# Initialize with memory enabled
agents = AgentTeam(
    agents=[agent],
    tasks=[],
    memory=True
)

# Search with reranking for better relevance
if agents.shared_memory:
    results = agents.shared_memory.search_long_term(
        "quantum computing applications",
        limit=10,
        rerank=True,  # Enable reranking
        relevance_cutoff=0.7  # Minimum relevance score
    )

Reranking Features

1. Semantic Reranking: Re-scores results based on semantic similarity
2. Context-Aware Ranking: Considers current context when ranking
3. Quality-Weighted Ranking: Combines relevance with quality scores

# Advanced reranking with multiple criteria
results = agents.shared_memory.search_long_term(
    query="machine learning frameworks",
    limit=20,  # Retrieve more initially
    rerank=True,  # Enable reranking
    relevance_cutoff=0.6,  # Minimum relevance
    rerank_config={
        "method": "semantic",  # or "hybrid", "quality-weighted"
        "top_k": 5,  # Return top 5 after reranking
        "consider_recency": True,  # Factor in timestamp
        "boost_user_memories": True  # Prioritize user-specific memories
    }
)

Custom Reranking Logic

Implement custom reranking strategies:

from datetime import datetime
from praisonaiagents import RerankStrategy

class DomainSpecificReranker(RerankStrategy):
    def rerank(self, results, query, context=None):
        """Custom reranking based on domain knowledge"""
        scored_results = []
        
        for result in results:
            score = result['relevance']
            
            # Boost technical content
            if 'technical' in result.get('metadata', {}).get('tags', []):
                score *= 1.2
                
            # Boost recent memories
            if result.get('timestamp'):
                days_old = (datetime.now() - result['timestamp']).days
                recency_boost = 1.0 + (0.1 * max(0, 30 - days_old) / 30)
                score *= recency_boost
                
            scored_results.append((score, result))
        
        # Sort by score and return top results
        scored_results.sort(key=lambda x: x[0], reverse=True)
        return [result for score, result in scored_results]

# Use custom reranker
# Note: Assuming 'agents' is defined from previous examples
if agents.shared_memory:
    agents.shared_memory.set_rerank_strategy(DomainSpecificReranker())

Reranking Performance

Optimize reranking for large result sets:

# Efficient two-stage retrieval and reranking
# Stage 1: Fast retrieval of candidates
candidates = agents.shared_memory.search_long_term(
    query="data analysis techniques",
    limit=50,  # Get more candidates
    rerank=False  # Skip reranking in first stage
)

# Stage 2: Precise reranking of top candidates
if len(candidates) > 10:
    reranked = agents.shared_memory.rerank_results(
        results=candidates[:20],  # Rerank top 20
        query="data analysis techniques",
        method="hybrid",
        relevance_cutoff=0.8
    )
else:
    reranked = candidates

Performance Optimisation

Batch Operations

# Batch store memories
memories = [
    ("Fact 1", 0.9),
    ("Fact 2", 0.85),
    ("Fact 3", 0.95)
]

for text, quality in memories:
    memory.store_long_term(
        text=text,
        quality=quality,
        memory={"user_id": "user123"}
    )

Complete Example

Personal Assistant with Memory

from praisonaiagents import Agent, Task, AgentTeam
from praisonaiagents import Memory

# Configure memory with quality scoring
memory_config = {
    "provider": "rag",
    "use_embedding": True,
    "embedding_model": "text-embedding-3-small"
}

# Create memory instance
memory = Memory(memory_config)

# Create assistant agent
assistant = Agent(
    name="Personal Assistant",
    instructions="""You are a personal assistant with perfect memory.
    Remember user preferences, important information, and context.
    Always use your memory to provide personalised responses.""",
    memory={"user_id": "john_doe"}
)

# Create task with memory integration
task = Task(
    description="Help me plan my day based on my preferences",
    agent=assistant,
    expected_output="Personalised daily schedule"
)

# Run the system
agents = AgentTeam(
    agents=[assistant],
    tasks=[task],
    memory=memory
)

# Memory automatically:
# 1. Retrieves user preferences
# 2. Searches relevant past interactions
# 3. Builds context for the task
# 4. Stores the interaction with quality score
result = agents.start()

# Store user feedback as high-quality memory
memory.store_long_term(
    text="User preferred morning schedule with exercise first",
    memory={"user_id": "john_doe"},
    completeness=0.9,
    relevance=1.0,
    clarity=0.95,
    accuracy=1.0
)

Best Practices

• Set minimum quality for critical information (0.8+)
• Use lower thresholds for exploratory searches
• Regular quality audits

• Periodically review and clean old memories
• Update quality scores as information ages
• Deduplicate similar memories

• Use user_id for personalisation
• Use agent_id for agent-specific knowledge
• Use run_id for session isolation

• Enable embeddings for semantic search
• Use quality filters to reduce search space
• Implement caching for frequent queries

Next Steps

Knowledge

Integrate with document knowledge bases

Sessions

Learn about stateful sessions