Schemas and profiles

A schema is a Pydantic model for your memories. With enable_inserts=False the manager keeps exactly one of it, a profile that updates field by field.

Example

from pydantic import BaseModel
from langmem import create_memory_manager

from memory_model import MemoryModel


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

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

Example

profiler = create_memory_manager(MemoryModel(), schemas=[Profile], enable_inserts=False)
profile = profiler.invoke({"messages": [{"role": "user", "content": "My name is Asha, please email me."}]})
print(profile[0].content)

schemas=[Profile] gives the model a Profile tool with the model's fields, so it fills in fields instead of writing sentences. The docstring becomes the tool's description. enable_inserts=False means: once a profile exists, update it, never add a second one.

Example

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

The same profile, same id, with city filled in by a JSON Patch on /city. The name and contact stayed. Your app can read profile.contact directly, no search needed.

Other schemas

Example

class OrderIssue(BaseModel):
    """A problem a customer had with an order."""

    order_id: str
    problem: str
    resolved: bool = False


manager = create_memory_manager("anthropic:claude-sonnet-4-5", schemas=[OrderIssue, Profile])

Several schemas give the model several tools. With inserts enabled, OrderIssue would collect many records while typed fields keep them consistent. LangMem's docs use the same idea for episodes: a schema with the situation, the reasoning and the result of an interaction that went well.

Try it yourself

Add order_ids: list[str] = [] to Profile and extract.
Set enable_inserts=True and send two separate names.
Print Profile.model_json_schema(): that is the tool the model sees.

PreviousUpdating and deleting memories

Slow is fine. Stopping is the only problem.