agno-memory

name: agno-memory description: | Add persistent memory to Agno agents. Covers MemoryManager, agentic memory, shared memory between agents, multi-user sessions, and memory tools. Trigger this skill when: importing agno.memory, configuring MemoryManager, enabling agentic memory, or asking "how do I add memory to an agent?" license: Apache-2.0 metadata: version: "1.0.0" author: agno-team tags: ["memory", "persistence", "agno", "user-preferences"]

Add Memory to Agno Agents

Memory lets agents remember user-level facts (preferences, goals, context) across all sessions. Different from storage (conversation history).

Quick Start

from agno.agent import Agent
from agno.db.sqlite import SqliteDb
from agno.memory import MemoryManager

db = SqliteDb(db_file="tmp/agents.db")

memory_manager = MemoryManager(
    model="openai:gpt-4o",
    db=db,
)

agent = Agent(
    name="Memory Agent",
    model="openai:gpt-4o",
    db=db,
    memory_manager=memory_manager,
    enable_agentic_memory=True,
    add_history_to_context=True,
    num_history_runs=5,
)

agent.print_response(
    "I prefer Python over JavaScript",
    user_id="user@example.com",
    stream=True,
)

# Retrieve stored memories
memories = agent.get_user_memories(user_id="user@example.com")
for m in memories:
    print(f"- {m.memory}")

Memory vs Storage

MemoryManager Configuration

from agno.memory import MemoryManager

memory_manager = MemoryManager(
    model="openai:gpt-4o",         # Model for memory extraction
    db=db,                          # Database to store memories
    additional_instructions="Capture user preferences and technical context.",
)

Memory Modes

Agentic Memory (Recommended)

The agent decides when to store/recall memories. More natural, fewer unnecessary writes:

agent = Agent(
    model="openai:gpt-4o",
    memory_manager=memory_manager,
    enable_agentic_memory=True,     # Agent chooses when to use memory
    db=db,
)

Always-On Memory

Run memory manager on every interaction. Guaranteed capture but slower:

agent = Agent(
    model="openai:gpt-4o",
    memory_manager=memory_manager,
    update_memory_on_run=True,      # Always extract memories
    db=db,
)

Memory with Custom Instructions

Guide what the memory manager captures:

memory_manager = MemoryManager(
    model="openai:gpt-4o",
    db=db,
    additional_instructions=[
        "Capture the user's programming language preferences.",
        "Remember project names and tech stacks.",
        "Track dietary restrictions and food preferences.",
        "Ignore small talk and greetings.",
    ],
)

Shared Memory Between Agents

Multiple agents can share the same memory store via the same db and user_id:

db = SqliteDb(db_file="tmp/shared.db")

memory_manager = MemoryManager(model="openai:gpt-4o", db=db)

assistant = Agent(
    name="Assistant",
    model="openai:gpt-4o",
    memory_manager=memory_manager,
    enable_agentic_memory=True,
    db=db,
)

researcher = Agent(
    name="Researcher",
    model="openai:gpt-4o",
    memory_manager=memory_manager,
    enable_agentic_memory=True,
    db=db,
)

# Both agents access the same user memories
assistant.print_response("I work on the Acme project", user_id="alice", stream=True)
researcher.print_response("What project does the user work on?", user_id="alice", stream=True)

Multi-User Sessions

Use user_id to isolate memories per user:

agent.print_response("I like dark mode", user_id="alice", stream=True)
agent.print_response("I like light mode", user_id="bob", stream=True)

alice_memories = agent.get_user_memories(user_id="alice")
bob_memories = agent.get_user_memories(user_id="bob")

Memory Tools

Give agents explicit tools to manage memories:

from agno.tools.memory import MemoryTools

agent = Agent(
    model="openai:gpt-4o",
    tools=[MemoryTools()],
    db=db,
    memory_manager=memory_manager,
)

Retrieving and Managing Memories

# Get all memories for a user
memories = agent.get_user_memories(user_id="alice")

# Memories have structure
for m in memories:
    print(f"ID: {m.memory_id}")
    print(f"Content: {m.memory}")
    print(f"Topics: {m.topics}")

Anti-Patterns

• Don't confuse memory with storage — memory is user-level facts, storage is chat history
• Don't forget user_id — without it, memories can't be isolated per user
• Don't use update_memory_on_run=True with enable_agentic_memory=True — pick one mode
• Don't skip db= — memories need a database backend to persist
• Don't store sensitive data — memories are plain text; avoid PII or secrets

Further Reading

For custom memory managers and advanced patterns, read references/api-patterns.md.