Profiles

A profile is a memory with a fixed shape: one Pydantic document per user, such as a name, a contact method and a city, that the manager keeps up to date field by field instead of adding new memories.

Last updated: 30 Sep, 2026 · LangMem 0.0.30

Some facts have exactly one current value: Asha has one name and one preferred contact method. For those, a profile is easier to use than a collection, because your code can read profile.contact directly, with no search.

Syntax:

manager = create_memory_manager(model, schemas=[Profile], enable_inserts=False)

The profile schema

A schema is a Pydantic model. Its docstring becomes the tool's description and its fields the tool's arguments, so the model fills in fields instead of writing sentences.

from pydantic import BaseModel


class Profile(BaseModel):
    """What we know about a support customer."""

    name: str | None = None
    contact: str | None = None
    city: str | None = None

A manager that keeps one profile

schemas=[Profile] replaces the default Memory tool with a Profile tool. enable_inserts=False means: once a profile exists, only update it, never add a second one.

profiler = create_memory_manager(
    model,
    schemas=[Profile],
    instructions="Fill in the customer's profile.",
    enable_inserts=False,
)

Filling in and updating Asha's profile

from langchain.chat_models import init_chat_model
from langmem import create_memory_manager

from pydantic import BaseModel


class Profile(BaseModel):
    """What we know about a support customer."""

    name: str | None = None
    contact: str | None = None
    city: str | None = None


model = init_chat_model("groq:openai/gpt-oss-120b", temperature=0)
profiler = create_memory_manager(
    model,
    schemas=[Profile],
    instructions="Fill in the customer's profile.",
    enable_inserts=False,
)

first = profiler.invoke({"messages": [{"role": "user", "content": "My name is Asha, please email me."}]})
print(first[0].content)

update = profiler.invoke({
    "messages": [{"role": "user", "content": "I moved to Pune last month."}],
    "existing": [(first[0].id, first[0].content)],
})
print(update[0].content, "| same id:", update[0].id == first[0].id)

name='Asha' contact=None city=None
name='Asha' contact=None city='Pune' | same id: True

What the two calls did

• First call: name filled, contact left empty. Asha said "please email me", yet contact is None. The model missed a field it could have filled; a clearer field description or instruction, such as "contact is the channel: email, phone or text", gives it a better chance.
• Second call: city filled in, same id. same id: True means the model patched the one profile instead of adding a second, because enable_inserts=False allows only updates once a profile exists.
• The name stayed. A patch changes only the fields it names; name was untouched.

Profile vs collection

When a profile fits

• Settings a user has one of: preferred name, language, timezone, contact method.
• Data your code needs as a value, such as the channel to send a notification on.
• A screen where users see and correct what the assistant knows about them.

Related

• Previous: Updates and deletes
• Next: Episodic memory
• Reference: How to manage user profiles