Sessions and memory

Persistent multi-turn conversations and per-user memory.

An agent needs two kinds of state: what was said in this thread, and what the agent knows about this user. Agno gives you both. You do not build a conversation store or a retrieval layer.

from agno.agent import Agent
from agno.db.postgres import PostgresDb
from agno.models.openai import OpenAIResponses

db = PostgresDb(db_url="postgresql+psycopg://ai:ai@localhost:5532/ai")

agent = Agent(
    model=OpenAIResponses(id="gpt-5.5"),
    db=db,
    add_history_to_context=True,
    num_history_runs=5,
    enable_agentic_memory=True,
)

agent.run(
    "My name is Sarah and I prefer email over phone.",
    user_id="[e-mail]",
    session_id="thread-42",
)
reply = agent.run(
    "What's the best way to reach me?",
    user_id="[e-mail]",
    session_id="thread-42",
).content
# Uses both the thread history and the stored memory about Sarah.

Sessions vs memory

They solve different problems. Use both.

	Session history	Memory
Stores	The messages in this conversation thread	Learned facts about the user
Scope	One session_id	One user_id, across all their sessions
Enable with	add_history_to_context=True	enable_agentic_memory=True or update_memory_on_run=True
Answers	”What did we just discuss?"	"What do I know about this person?”

Identifiers

Identifier	Distinguishes	Maps to in your product
user_id	The person	Your auth subject (user ID, email)
session_id	A conversation thread for that person	A chat tab, a Slack thread, a support case

Pass both on every run. Threads are scoped by session_id. Memory is scoped by user_id.

Memory: automatic or agentic

Mode	Set	Use when
Automatic	update_memory_on_run=True	Compute memory after every run.
Agentic	enable_agentic_memory=True	The agent decides using tool calls.

Reading memory back

For a profile screen or a debug view, pull a user’s memories directly.

memories = agent.get_user_memories(user_id="[e-mail]")

Long conversations

History grows. Two ways to keep token cost bounded without losing continuity:

Technique	Effect
num_history_runs=N	Only the last N turns flow into context
Session summaries (enable_session_summaries=True)	Older turns are condensed into a running summary

Next steps

Task	Guide
Put this behind an HTTP API	Serve as an API
Carry memory across Slack and web	Interfaces
Give the agent live data	Connecting your data

Developer Resources

• Sessions
• Session management
• Memory
• Memory cookbook