Memory

The Contacts agent is a mini CRM that remembers people, the facts you tell it about them, who knows whom, and what’s happening between them. Three memory stores, each scoped to a different shape of state:

from agno.learn import (
    LearningMachine, LearningMode,
    UserProfileConfig, EntityMemoryConfig, SessionContextConfig,
)

contacts = Agent(
    id="contacts",
    name="Contacts",
    model=MODEL,
    db=agent_db,
    learning=LearningMachine(
        user_profile=UserProfileConfig(mode=LearningMode.ALWAYS),
        entity_memory=EntityMemoryConfig(
            mode=LearningMode.AGENTIC,
            enable_create_entity=True,
            enable_add_fact=True,
            enable_add_relationship=True,
            enable_add_event=True,
        ),
        session_context=SessionContextConfig(
            mode=LearningMode.ALWAYS,
            enable_planning=True,
        ),
    ),
    enable_agentic_memory=True,
)

Three stores, three shapes

Store	Cardinality	Mode	Example
User profile	One record per user	`ALWAYS` (auto-extracted)	“Prefers async communication, based in PT, role: PM”
Entity memory	Graph of entities + facts + relationships + events	`AGENTIC` (agent writes)	Sarah Chen → role: VP Eng → company: Acme → “joined 2024-Q4”
Session context	Per session	`ALWAYS`	”User is planning a Q1 offsite, talking through agenda”

LearningMode.ALWAYS runs an extractor on every turn. LearningMode.AGENTIC gives the agent tools to write when it judges something worth keeping.

Entity memory in detail

The Contacts agent gets four entity-memory tools when configured as above:

Tool	What it does
`create_entity(name, type)`	Adds a new entity (Person, Company, Project, …)
`add_fact(entity, fact)`	Attaches a fact to an entity
`add_relationship(entity_a, entity_b, type)`	Links two entities
`add_event(entity, event, when)`	Records a time-bound event

Example chat that builds up the graph:

User: Sarah Chen joined Acme as VP Engineering last month.
Contacts: → create_entity("Sarah Chen", "Person")
          → add_fact("Sarah Chen", "VP Engineering at Acme")
          → add_relationship("Sarah Chen", "Acme", "employed_by")
          → add_event("Sarah Chen", "joined Acme as VP Eng", "2026-03-15")

Next session, “what do I know about Sarah?” returns the full picture without the user re-telling.

User profile

User profile is one record per user_id, updated in place. Whatever the agent learns about you (your role, your preferences, your working style) goes here.

User: I prefer concise responses, no preamble.
Contacts: → user_profile updated: { communication_style: "concise, direct" }

Future responses across any session, any agent that shares the same db, get terser automatically.

Session context

Session context is plan-shaped. “What’s this conversation actually about? What does the user want by the end?” The agent uses it to keep multi-turn plans coherent.

Session goal: Plan Q1 offsite agenda
Open items: pick venue, set theme, draft schedule
Decisions so far: 3 days, 25 people, focus on roadmap alignment

Useful for long, exploratory sessions where the user pivots and the agent needs to stay anchored.

Vanilla agentic memory vs LearningMachine

For most agents, enable_agentic_memory=True is enough: a single bucket of facts that the agent writes to via tools. LearningMachine is right when the shape of memory matters:

A relationship graph (Contacts) needs entities and edges, not flat strings.
A self-improving SQL agent (Dash) needs to separate “validated queries” from “error patterns I’ve fixed”.

When in doubt, start with enable_agentic_memory=True and move to LearningMachine once flat memory feels too lossy.

See it in action

@Contacts Sarah Chen joined Acme as VP Eng last month
@Contacts what do I know about Sarah?
@Contacts who at Acme have I been talking to?
@Contacts I'm meeting Sarah tomorrow about the platform roadmap
@Contacts later: what was the meeting with Sarah about?

Each turn updates the entity graph. The agent uses the graph on every retrieval. Source: agents/contacts/agent.py, Learning Machine docs

Learning →

Welcome

Tutorials

Runtime

Demo OS

Three stores, three shapes

Entity memory in detail

User profile

Session context

Vanilla agentic memory vs LearningMachine

See it in action

Next

​Three stores, three shapes

​Entity memory in detail

​User profile

​Session context

​Vanilla agentic memory vs LearningMachine

​See it in action

​Next

Three stores, three shapes

Entity memory in detail

User profile

Session context

Vanilla agentic memory vs LearningMachine

See it in action

Next