Stockage de la mémoire à long terme entre threads dans LangGraph

Pourquoi la mémoire de fil à court terme n’est pas suffisante

L’historique de chat sauvegardé couvre un seul fil. Les utilisateurs réels passent d’un fil à l’autre — débogage de LangChain le lundi, planification de voyages mardi — ce qui fait que leurs informations sont dispersées. La mémoire à long terme (LTM) offre un stockage durable : un extracteur enregistre les détails personnels à partir des échanges, puis l’assistant consulte ce stockage avant de répondre aux questions. Si vous préférez Python une fois et demandez ensuite les nombres de Fibonacci, l’assistant peut réutiliser cette préférence sans avoir à être informé à nouveau.

Les stockages en tant que surface mémoire de LangGraph

LangGraph gère la LTM via des stockages. BaseStore définit les opérations de création, de recherche, d’édition et de suppression. Les backends courants sont : InMemoryStore pour les prototypes en RAM ; PostgresStore et RedisStore lorsque les redémarrages ne doivent pas effacer la mémoire.

Opérations sur les stockages indépendants

Traitez les noms d’espace comme des dossiers. Ce sont des tuples Python — "users", "U1" ou, de manière plus profonde, "users", "U1", "profile".

Créez-les avec put(namespace, key, values). Récupérez un élément avec get lorsque la clé est connue ; listez un nom d’espace avec search.

from langgraph.store.memory import InMemoryStore

# creating a store
store = InMemoryStore()

# creating a namespace
namespace = ("user", "u1")

Création de mémoires

# Adding memories
store.put(namespace, "1", {"data": "User likes pizza"})
store.put(namespace, "2", {"data": "User prefers dark mode"})

# another namespace
namespace2 = ("user", "u2")

# Add memories
store.put(namespace2, "1", {"data": "User likes pasta"})
store.put(namespace2, "2", {"data": "User prefers grid style navigation"})

Récupération d’une mémoire

# store.get(namespace, key)
store.get(namespace2, "2")

Item(namespace=['user', 'u2'], key='2', value={'data': 'User prefers grid style navigation'}, created_at='2026-01-06T15:05:16.376069+00:00', updated_at='2026-01-06T15:05:16.376070+00:00')

Récupération de tout dans un nom d’espace

items = store.search(namespace2)

for item in items:
    print(item.value)

{'data': 'User likes pasta'}
{'data': 'User prefers grid style navigation'}

Recherche sémantique lorsque le ensemble grandit

En fournissant une centaine de données de profil dans la requête, on surcharge le modèle. Indexez le stockage à l’aide d’un modèle d’incorporation (par exemple text-embedding-3-small) afin que la fonction search puisse accepter une query et un limit, ne retournant que des vecteurs proches de la question posée.

Semantic Search
from langchain_openai import OpenAIEmbeddings
from dotenv import load_dotenv

load_dotenv()

embedding_model = OpenAIEmbeddings(model='text-embedding-3-small')
store = InMemoryStore(index={'embed': embedding_model, 'dims':1536})
namespace = ('users', 'u1')
store.put(namespace, "1", {"data": "User prefers concise answers over long explanations"})
store.put(namespace, "2", {"data": "User likes examples in Python"})
store.put(namespace, "3", {"data": "User usually works late at night"})
store.put(namespace, "4", {"data": "User prefers dark mode in applications"})
store.put(namespace, "5", {"data": "User is learning machine learning"})
store.put(namespace, "6", {"data": "User dislikes overly theoretical explanations"})
store.put(namespace, "7", {"data": "User prefers step-by-step reasoning"})
store.put(namespace, "8", {"data": "User is based in India"})
store.put(namespace, "9", {"data": "User likes real-world analogies"})
store.put(namespace, "10", {"data": "User prefers bullet points over paragraphs"})
items = store.search(namespace, query="what is the user currently learning", limit=1)

for item in items:
    print(item.value)
# output :{'data': 'User is learning machine learning'}
items = store.search(namespace, query="what are user's preferences", limit=3)

for item in items:
    print(item.value)
# output :
# {'data': 'User prefers dark mode in applications'}
# {'data': 'User prefers bullet points over paragraphs'}
# {'data': 'User prefers concise answers over long explanations'}

Personnalisation en lecture seule dans un nœud de chat

Connectez le stockage à un nœud de chat LangGraph qui reçoit state, config (identifiant de l’utilisateur ou du fil de discussion) et store. Construisez l’espace de noms à partir de l’identifiant de l’utilisateur, recherchez les mémoires correspondantes, intégrez-les dans la prompt système, et indiquez au modèle d’utiliser des noms, des projets et un ton connus, ainsi que quelques questions complémentaires pertinentes.

Chatbot Reading Existing Memories
from dotenv import load_dotenv
load_dotenv()

from langchain_openai import ChatOpenAI
from langchain_core.runnables import RunnableConfig
from langchain_core.messages import SystemMessage

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.store.memory import InMemoryStore
from langgraph.store.base import BaseStore
# ----------------------------
# 1) Create LTM store + seed memories (done BEFORE running the graph)
# ----------------------------
store = InMemoryStore()

user_id = "u1"

# Store user details as a single blob (simple for teaching)
# You can also split into multiple records; this keeps it easy.
user_details = ("user", user_id, "details")

store.put(user_details, "profile_1", {"data": "Name: Nitish"})
store.put(user_details, "profile_2", {"data": "Profession: Teaches AI on YouTube"})
store.put(user_details, "preference_1", {"data": "Prefers concise answers"})
store.put(user_details, "preference_2", {"data": "Likes examples in Python"})
store.put(user_details, "project_1", {"data": "Building MCP servers (Python-based project)"})
# ----------------------------
# 2) System prompt template (your prompt)
# ----------------------------
SYSTEM_PROMPT_TEMPLATE = """You are a helpful assistant with memory capabilities.
If user-specific memory is available, use it to personalize
your responses based on what you know about the user.

Your goal is to provide relevant, friendly, and tailored
assistance that reflects the user’s preferences, context, and past interactions.

If the user’s name or relevant personal context is available, always personalize your responses by:
    – Always Address the user by name (e.g., "Sure, Nitish...") when appropriate
    – Referencing known projects, tools, or preferences (e.g., "your MCP  server python based project")
    – Adjusting the tone to feel friendly, natural, and directly aimed at the user

Avoid generic phrasing when personalization is possible. For example, instead of "In TypeScript apps..."
say "Since your project is built with TypeScript..."

Use personalization especially in:
    – Greetings and transitions
    – Help or guidance tailored to tools and frameworks the user uses
    – Follow-up messages that continue from past context

Always ensure that personalization is based only on known user details and not assumed.

In the end suggest 3 relevant further questions based on the current response and user profile

The user’s memory (which may be empty) is provided as: {user_details_content}
"""
# ----------------------------
# 3) Build graph: START -> chat -> END (read-only LTM)
# ----------------------------
llm = ChatOpenAI(model="gpt-4o-mini")
def chat_node(state: MessagesState, config: RunnableConfig, store: BaseStore):

    user_id = config["configurable"]["user_id"]

    # Read-only: fetch user details memory (no writes)
    user_details = ("user", user_id, "details")
    items = store.search(user_details)

    # Convert memory items into a string blob for {user_details_content}
    # Keep it dead simple for teaching.
    if items:
        user_details_content = "\n".join(f"- {it.value.get('data', '')}" for it in items)
    else:
        user_details_content = ""  # prompt says it may be empty

    system_prompt = SYSTEM_PROMPT_TEMPLATE.format(
        user_details_content=user_details_content
    )

    system_msg = SystemMessage(content=system_prompt)

    response = llm.invoke([system_msg] + state["messages"])
    return {"messages": [response]}
builder = StateGraph(MessagesState)
builder.add_node("chat", chat_node)
builder.add_edge(START, "chat")
builder.add_edge("chat", END)

graph = builder.compile(store=store)

graph

# ----------------------------
# 4) Run it (provide user_id in config)
# ----------------------------
config = {"configurable": {"user_id": "u1"}}

result = graph.invoke(
    {"messages": [{"role": "user", "content": "Explain gen ai in simple terms."}]},
    config,
)

print(result["messages"][-1].content)
# Sure, Nitish! Generative AI, or Gen AI, refers to a type of artificial intelligence that creates new content or data instead of just analyzing existing information. It can generate text, images, music, and more by learning patterns from large sets of data. For example, it can write a story, design an artwork, or compose a song based on the styles of various creators it has learned from.

#If you're considering using examples or applications of Gen AI for your YouTube teaching, there are plenty of interesting case studies you could explore, especially in creative fields.

#Here are a few questions to consider:

#1. Would you like to see specific examples of Gen AI in Python?
#2. Are you interested in exploring tools or libraries for building Gen AI projects?
#3. How would you like to incorporate Gen AI concepts into your teaching materials?

Création de nouvelles mémoires pendant la conversation

Chatbot Creating New Memories
from dotenv import load_dotenv
load_dotenv()

import uuid
from typing import List
from pydantic import BaseModel, Field

from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage
from langchain_core.runnables import RunnableConfig

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.store.memory import InMemoryStore
from langgraph.store.base import BaseStore
# ----------------------------
# 1) LTM store
# ----------------------------
store = InMemoryStore()
# ----------------------------
# 2) LLM that decides what to remember (structured output)
# ----------------------------
extractor_llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
class MemoryDecision(BaseModel):
    should_write: bool = Field(description="Whether to store any memories")
    memories: List[str] = Field(default_factory=list, description="Atomic user memories to store")
memory_extractor = extractor_llm.with_structured_output(MemoryDecision)
# ----------------------------
# 3) Graph: START -> remember -> END
#    (Creates memories, but does NOT use them to answer)
# ----------------------------
def remember_only_node(state: MessagesState, config: RunnableConfig, store: BaseStore):

    user_id = config["configurable"]["user_id"]

    namespace = ("user", user_id, "details")

    # take latest user message
    last_msg = state["messages"][-1].content

    # LLM decides what to store
    decision: MemoryDecision = memory_extractor.invoke(
        [
            SystemMessage(
                content=(
                    "Extract LONG-TERM memories from the user's message.\n"
                    "Only store stable, user-specific info (identity, preferences, ongoing projects).\n"
                    "Do NOT store transient info.\n"
                    "Return should_write=false if nothing is worth storing.\n"
                    "Each memory should be a short atomic sentence."
                )
            ),
            {"role": "user", "content": last_msg},
        ]
    )

    # Write to store (LTM)
    if decision.should_write:
        for mem in decision.memories:
            store.put(namespace, str(uuid.uuid4()), {"data": mem})

    # IMPORTANT: we are NOT using memory, not even responding with the LLM.
    # We just return a fixed acknowledgement.
    return {"messages": [{"role": "assistant", "content": "Noted."}]}
builder = StateGraph(MessagesState)
builder.add_node("remember", remember_only_node)
builder.add_edge(START, "remember")
builder.add_edge("remember", END)

graph = builder.compile(store=store)

graph

# ----------------------------
# 4) Demo
# ----------------------------
config = {"configurable": {"user_id": "u1"}}

res = graph.invoke({"messages": [{"role": "user", "content": "Hi my name is nitish"}]},config)
print("Assistant:", res["messages"][-1].content)
#output : Assistant: Noted.
res = graph.invoke({"messages": [{"role": "user", "content": "I teach AI on youtube"}]},config)
print("Assistant:", res["messages"][-1].content)
#output :Assistant: Noted.
res = graph.invoke({"messages": [{"role": "user", "content": "My favorite programming language is Python"}]},config)
print("Assistant:", res["messages"][-1].content)
# output :Assistant: Noted.
items = store.search(("user", "u1", "details"))

for item in items:
    print(item.value['data'])
#output:
#  User's name is Nitish.
#  User teaches AI on YouTube.
 # User's favorite programming language is Python.
  #User's name is Nitish.
#  User teaches AI on YouTube.
#  User's favorite programming language is Python.

Les écritures automatiques nécessitent un nœud de mémorisation avant la conversation. Un modèle d’extraction dédié génère une sortie structurée MemoryDecision indiquant s’il faut écrire et quels éléments. Sans suppression des doublons, la répétition de « mon nom est … » crée des clones. Mettez à jour les éléments en utilisant un drapeau is_new par rapport aux mémoires existantes et conservez uniquement les nouvelles.

Chatbot Creating New Memories(without Duplication)
from dotenv import load_dotenv
load_dotenv()

import uuid
from typing import List
from pydantic import BaseModel, Field

from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage
from langchain_core.runnables import RunnableConfig

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.store.memory import InMemoryStore
from langgraph.store.base import BaseStore
# ----------------------------
# 1) LTM store
# ----------------------------
store = InMemoryStore()
# ----------------------------
# 2) LLMs
#    - memory_llm: extracts candidate memories + tells if each is NEW (no duplicate_of needed)
# ----------------------------
memory_llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
class MemoryItem(BaseModel):
    text: str = Field(description="Atomic user memory as a short sentence")
    is_new: bool = Field(description="True if this memory is NEW and should be stored. False if duplicate/already known.")
class MemoryDecision(BaseModel):
    should_write: bool = Field(description="Whether to store any memories")
    memories: List[MemoryItem] = Field(default_factory=list, description="Atomic user memories to store")
memory_extractor = memory_llm.with_structured_output(MemoryDecision)
MEMORY_PROMPT = """You are responsible for updating and maintaining accurate user memory.

CURRENT USER DETAILS (existing memories):
{user_details_content}

TASK:
- Review the user's latest message.
- Extract user-specific info worth storing long-term (identity, stable preferences, ongoing projects/goals).
- For each extracted item, set is_new=true ONLY if it adds NEW information compared to CURRENT USER DETAILS.
- If it is basically the same meaning as something already present, set is_new=false.
- Keep each memory as a short atomic sentence.
- No speculation; only facts stated by the user.
- If there is nothing memory-worthy, return an empty list.
"""
def chat_creates_memory_node(state: MessagesState, config: RunnableConfig, store: BaseStore):

    user_id = config["configurable"]["user_id"]

    namespace = ("user", user_id, "details")

    # A) Load existing memories
    existing_items = store.search(namespace)
    existing_texts = [it.value.get("data", "") for it in existing_items if it.value.get("data")]
    user_details_content = "\n".join(f"- {t}" for t in existing_texts) if existing_texts else "(empty)"

    # B) Latest user message
    last_text = state["messages"][-1]

    # C) LLM extracts memories + marks new vs duplicate
    decision: MemoryDecision = memory_extractor.invoke(
        [
            SystemMessage(content=MEMORY_PROMPT.format(user_details_content=user_details_content)),
            {"role": "user", "content": f"USER MESSAGE:\n{last_text}"},
        ]
    )

    # D) Store ONLY new memories
    if decision.should_write:
        for mem in decision.memories:
            if mem.is_new:
                store.put(namespace, str(uuid.uuid4()), {"data": mem.text})

    return {"messages": [{"role": "assistant", "content": "Noted."}]}
# ----------------------------
# 4) Build graph: START -> chat -> END
# ----------------------------
builder = StateGraph(MessagesState)
builder.add_node("chat", chat_creates_memory_node)
builder.add_edge(START, "chat")
builder.add_edge("chat", END)

graph = builder.compile(store=store)
config = {"configurable": {"user_id": "u1"}}
r1 = graph.invoke({"messages": [{"role": "user", "content": "My name is Nitish"}]}, config)
print("Assistant:", r1["messages"][-1].content)
# output: Assistant: Noted.
# Likely duplicate of the first memory -> is_new should be false -> saved=0
r2 = graph.invoke({"messages": [{"role": "user", "content": "I like Python for programming."}]}, config)
print("\nAssistant:", r2["messages"][-1].content)
# output:Assistant: Noted.
for it in store.search(("user", "u1", "details")):
    print(it.value['data'])
# output:The user's name is Nitish.
# output:Nitish likes Python for programming.

Graphe fusionné et persistance en production

Structure finale : début → mémorisation (extraction et enregistrement) → conversation (lecture et réponse) → fin. L’agent apprend au cours de la conversation et utilise immédiatement les nouvelles informations.

Merged Workflow
from dotenv import load_dotenv
load_dotenv()

import uuid
from typing import List
from pydantic import BaseModel, Field

from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage
from langchain_core.runnables import RunnableConfig

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.store.memory import InMemoryStore
from langgraph.store.base import BaseStore
# ----------------------------
# 1) LTM store (START EMPTY)
# ----------------------------
store = InMemoryStore()
# ----------------------------
# 2) System prompt
# ----------------------------
SYSTEM_PROMPT_TEMPLATE = """You are a helpful assistant with memory capabilities.
If user-specific memory is available, use it to personalize
your responses based on what you know about the user.

Your goal is to provide relevant, friendly, and tailored
assistance that reflects the user’s preferences, context, and past interactions.

If the user’s name or relevant personal context is available, always personalize your responses by:
    – Always Address the user by name (e.g., "Sure, Nitish...") when appropriate
    – Referencing known projects, tools, or preferences (e.g., "your MCP server python based project")
    – Adjusting the tone to feel friendly, natural, and directly aimed at the user

Avoid generic phrasing when personalization is possible.

Use personalization especially in:
    – Greetings and transitions
    – Help or guidance tailored to tools and frameworks the user uses
    – Follow-up messages that continue from past context

Always ensure that personalization is based only on known user details and not assumed.

In the end suggest 3 relevant further questions based on the current response and user profile

The user’s memory (which may be empty) is provided as: {user_details_content}
"""
# ----------------------------
# 3) Memory extraction LLM
# ----------------------------
memory_llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
class MemoryItem(BaseModel):
    text: str = Field(description="Atomic user memory")
    is_new: bool = Field(description="True if new, false if duplicate")
class MemoryDecision(BaseModel):
    should_write: bool
    memories: List[MemoryItem] = Field(default_factory=list)
memory_extractor = memory_llm.with_structured_output(MemoryDecision)
MEMORY_PROMPT = """You are responsible for updating and maintaining accurate user memory.

CURRENT USER DETAILS (existing memories):
{user_details_content}

TASK:
- Review the user's latest message.
- Extract user-specific info worth storing long-term (identity, stable preferences, ongoing projects/goals).
- For each extracted item, set is_new=true ONLY if it adds NEW information compared to CURRENT USER DETAILS.
- If it is basically the same meaning as something already present, set is_new=false.
- Keep each memory as a short atomic sentence.
- No speculation; only facts stated by the user.
- If there is nothing memory-worthy, return should_write=false and an empty list.
"""
# ----------------------------
# 4) Node 1: remember
# ----------------------------
def remember_node(state: MessagesState, config: RunnableConfig, *, store: BaseStore):
    user_id = config["configurable"]["user_id"]
    ns = ("user", user_id, "details")

    # existing memory
    items = store.search(ns)
    existing = "\n".join(it.value["data"] for it in items) if items else "(empty)"

    # last user message
    last_msg = state["messages"][-1].content

    decision: MemoryDecision = memory_extractor.invoke(
        [
            SystemMessage(content=MEMORY_PROMPT.format(user_details_content=existing)),
            {"role": "user", "content": last_msg},
        ]
    )

    if decision.should_write:
        for mem in decision.memories:
            if mem.is_new:
                store.put(ns, str(uuid.uuid4()), {"data": mem.text})

    return {}  # no message change
# ----------------------------
# 5) Node 2: chat
# ----------------------------
chat_llm = ChatOpenAI(model="gpt-4o-mini")
def chat_node(state: MessagesState, config: RunnableConfig, *, store: BaseStore):
    user_id = config["configurable"]["user_id"]
    ns = ("user", user_id, "details")

    items = store.search(ns)
    user_details = "\n".join(it.value["data"] for it in items) if items else ""

    system_msg = SystemMessage(
        content=SYSTEM_PROMPT_TEMPLATE.format(
            user_details_content=user_details or "(empty)"
        )
    )

    response = chat_llm.invoke([system_msg] + state["messages"])
    return {"messages": [response]}
# ----------------------------
# 6) Graph
# ----------------------------
builder = StateGraph(MessagesState)
builder.add_node("remember", remember_node)
builder.add_node("chat", chat_node)

builder.add_edge(START, "remember")
builder.add_edge("remember", "chat")
builder.add_edge("chat", END)

graph = builder.compile(store=store)

graph

# ----------------------------
# 7) Demo
# ----------------------------
config = {"configurable": {"user_id": "u1"}}
result = graph.invoke({"messages": [{"role": "user", "content": "Hi, my name is Nitish"}]}, config)
result['messages'][-1].content
# output:'Hi Nitish! It’s great to meet you. How can I assist you today?'

for it in store.search(("user", "u1", "details")):
    print(it.value["data"])
result = graph.invoke({"messages": [{"role": "user", "content": "I teach AI on YouTube"}]}, config)
print(result['messages'][-1].content)
# output:That's great, Nitish! Teaching AI on YouTube must be really fulfilling. What topics do you cover in your videos? Are there any specific projects or concepts you find your audience is particularly interested in?
for it in store.search(("user", "u1", "details")):
    print(it.value["data"])
# output:User's name is Nitish.
# output:Nitish teaches AI on YouTube.
result = graph.invoke({"messages": [{"role": "user", "content": "Explain GenAI simply"}]}, config)
print(result['messages'][-1].content)
# output:Sure, Nitish! Generative AI, often referred to as GenAI, is a type of artificial intelligence that can create new content based on the data it's trained on. This includes generating text, images, music, and even video.

# output:Here's a simple breakdown:

# output:1. **Learning from Examples:** GenAI is trained on vast amounts of data, learning the patterns, styles, and structures of the content it sees.

# output:2. **Creating Content:** Once trained, it can generate new content that resembles what it learned. For example, it can write a story in a specific style, create artwork, or compose music.

# output:3. **Applications:** It's used in various fields, from helping writers brainstorm ideas to producing artwork and even customer service chatbots that can engage with users in a conversational manner.

# output:If you're planning to teach this topic on your YouTube channel, emphasizing practical examples and real-world applications could be very engaging for your audience!

Do you want to explore specific applications of GenAI, tools for using it, or perhaps the ethics involved?
for it in store.search(("user", "u1", "details")):
    print(it.value["data"])
# output:User's name is Nitish.
# output:Nitish teaches AI on YouTube.

InMemoryStore disparaît lorsque le processus s’arrête. Exécutez Postgres (souvent via Docker), installez les paquets de pointage/stockage de Postgres, et remplacez le stockage par une implémentation basée sur des connexions. Après un redémarrage complet, les requêtes vers la base de données doivent toujours restituer les informations du profil enregistré.

services:
  postgres:
    image: postgres:16
    environment:
      POSTGRES_USER: postgres
      POSTGRES_PASSWORD: postgres
      POSTGRES_DB: postgres
    ports:
      - "5442:5432"

! pip install -U "psycopg[binary,pool]" langgraph langgraph-checkpoint-postgres
from dotenv import load_dotenv
load_dotenv()

import uuid
from typing import List
from pydantic import BaseModel, Field

from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage
from langchain_core.runnables import RunnableConfig

from langgraph.graph import StateGraph, START, END, MessagesState
from langgraph.store.postgres import PostgresStore
from langgraph.store.base import BaseStore
# ----------------------------
# 2) System prompt
# ----------------------------
SYSTEM_PROMPT_TEMPLATE = """You are a helpful assistant with memory capabilities.
If user-specific memory is available, use it to personalize
your responses based on what you know about the user.

Your goal is to provide relevant, friendly, and tailored
assistance that reflects the user’s preferences, context, and past interactions.

If the user’s name or relevant personal context is available, always personalize your responses by:
    – Always Address the user by name (e.g., "Sure, Nitish...") when appropriate
    – Referencing known projects, tools, or preferences (e.g., "your MCP server python based project")
    – Adjusting the tone to feel friendly, natural, and directly aimed at the user

Avoid generic phrasing when personalization is possible.

Use personalization especially in:
    – Greetings and transitions
    – Help or guidance tailored to tools and frameworks the user uses
    – Follow-up messages that continue from past context

Always ensure that personalization is based only on known user details and not assumed.

In the end suggest 3 relevant further questions based on the current response and user profile

The user’s memory (which may be empty) is provided as: {user_details_content}
"""
# ----------------------------
# 3) Memory extraction LLM
# ----------------------------
memory_llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
class MemoryItem(BaseModel):
    text: str = Field(description="Atomic user memory")
    is_new: bool = Field(description="True if new, false if duplicate")
class MemoryDecision(BaseModel):
    should_write: bool
    memories: List[MemoryItem] = Field(default_factory=list)
memory_extractor = memory_llm.with_structured_output(MemoryDecision)
MEMORY_PROMPT = """You are responsible for updating and maintaining accurate user memory.

CURRENT USER DETAILS (existing memories):
{user_details_content}

TASK:
- Review the user's latest message.
- Extract user-specific info worth storing long-term (identity, stable preferences, ongoing projects/goals).
- For each extracted item, set is_new=true ONLY if it adds NEW information compared to CURRENT USER DETAILS.
- If it is basically the same meaning as something already present, set is_new=false.
- Keep each memory as a short atomic sentence.
- No speculation; only facts stated by the user.
- If there is nothing memory-worthy, return should_write=false and an empty list.
"""
# ----------------------------
# 3) Nodes
# ----------------------------
def remember_node(state: MessagesState, config: RunnableConfig, *, store: BaseStore):
    user_id = config["configurable"]["user_id"]
    ns = ("user", user_id, "details")

    # existing memory (all items under namespace)
    items = store.search(ns)
    existing = "\n".join(it.value.get("data", "") for it in items) if items else "(empty)"

    # latest user message
    last_text = state["messages"][-1].content

    decision: MemoryDecision = memory_extractor.invoke(
        [
            SystemMessage(content=MEMORY_PROMPT.format(user_details_content=existing)),
            {"role": "user", "content": last_text},
        ]
    )

    if decision.should_write:
        for mem in decision.memories:
            if mem.is_new and mem.text.strip():
                store.put(ns, str(uuid.uuid4()), {"data": mem.text.strip()})

    return {}
chat_llm = ChatOpenAI(model="gpt-4o-mini")
def chat_node(state: MessagesState, config: RunnableConfig, *, store: BaseStore):
    user_id = config["configurable"]["user_id"]
    ns = ("user", user_id, "details")

    items = store.search(ns)
    user_details = "\n".join(it.value.get("data", "") for it in items) if items else ""

    system_msg = SystemMessage(
        content=SYSTEM_PROMPT_TEMPLATE.format(user_details_content=user_details or "(empty)")
    )

    response = chat_llm.invoke([system_msg] + state["messages"])
    return {"messages": [response]}
# ----------------------------
# 4) Build graph
# ----------------------------
builder = StateGraph(MessagesState)
builder.add_node("remember", remember_node)
builder.add_node("chat", chat_node)
builder.add_edge(START, "remember")
builder.add_edge("remember", "chat")
builder.add_edge("chat", END)
<langgraph.graph.state.StateGraph at 0x110ce4a90>
# ----------------------------
# 5) Use PostgresStore (PERSISTENT LTM)
# ----------------------------
DB_URI = "postgresql://postgres:postgres@localhost:5442/postgres?sslmode=disable"

with PostgresStore.from_conn_string(DB_URI) as store:
    # IMPORTANT: run ONCE the first time you use this database
    store.setup()

    graph = builder.compile(store=store)

    config = {"configurable": {"user_id": "u1"}}

    graph.invoke({"messages": [{"role": "user", "content": "Hi, my name is Nitish"}]}, config)
    graph.invoke({"messages": [{"role": "user", "content": "I teach AI on YouTube"}]}, config)

    out = graph.invoke({"messages": [{"role": "user", "content": "Explain GenAI simply"}]}, config)
    print(out["messages"][-1].content)

    print("\n--- Stored Memories (from Postgres) ---")
    for it in store.search(("user", "u1", "details")):
        print(it.value["data"])
# output:Sure, Nitish! Generative AI, or GenAI, refers to a type of artificial intelligence that can create new content, like images, text, music, or even code, on its own. It learns from a large dataset and uses patterns it recognizes to generate something that resembles what it has seen before.

# output:For example, if you input a topic or a few words into a generative AI model, it can produce a coherent piece of writing or generate a unique piece of art. Popular examples include models like GPT-3 for text and DALL-E for images.

# output:The key point is that GenAI is designed to take what it has learned and create new, original outputs based on that learning. If you’re discussing this concept in your YouTube teachings, it might be interesting for your audience to see how GenAI can assist in creative processes or even in automating certain tasks.

# output:Would you like to dive deeper into a specific application of GenAI, or perhaps explore its implications for teaching AI?

# output:--- Stored Memories (from Postgres) ---
# output:Nitish teaches AI on YouTube.
# output:User's name is Nitish.
Check Persistence
from langgraph.store.postgres import PostgresStore

DB_URI = "postgresql://postgres:postgres@localhost:5442/postgres?sslmode=disable"

with PostgresStore.from_conn_string(DB_URI) as store:
    ns = ("user", "u1", "details")
    items = store.search(ns)

for it in items:
    print(it.value["data"])
# output: Nitish teaches AI on YouTube.
# output: User's name is Nitish.

Péripéties pratiques lors du câblage de LTM

Les espaces de noms semblent simples tant que deux fonctionnalités ne partagent pas ("users", user_id) et n’écrasent pas mutuellement leurs clés. Préférez des tuples ciblant spécifiquement chaque fonctionnalité, tels que ("users", user_id, "prefs") plutôt que ("users", user_id, "facts"), et documentez qui peut modifier chaque élément. Les clés doivent être des identifiants stables, et non des phrases de texte libre ; placez la phrase dans le contenu de la valeur.

La qualité de la recherche sémantique dépend du choix des embeddings ainsi que de la manière dont les informations sont formulées. Stocker « likes Python » puis interroger « preferred programming language » ne fonctionne que si l’espace d’embeddings regroupe ces expressions. Les re-embeddings périodiques après les mises à jour du modèle, bien que fastidieux, permettent d’éviter des dégradations silencieuses dans les résultats de recherche.

Les chemins d’écriture nécessitent des limites de débit. Un extracteur qui traite chaque énoncé de l’utilisateur peut inonder Postgres de quasi-dupliques, même après les vérifications is_new, en particulier lorsque le modèle est hésitant. Limitez le nombre d’écritures par tour, effectuez des résumés à la fin de chaque session en lot, et surveillez la croissance du stock par utilisateur. Pour le environnement de production, associez le stock à long terme à un point de contrôle à court terme afin que les dialogues locaux aux threads et la mémoire des profils entre threads restent conceptuellement séparés, même s’ils partagent l’infrastructure.

Hygiène des chemins de lecture

D’un point de vue lecture, ne chargez jamais l’ensemble de l’espace de noms dans la fenêtre système une fois que la personnalisation fonctionne dans les démos. Restreignez les souvenirs récupérés, indiquez quels clés d’espace de noms y ont contribué, et permettez à un opérateur de désactiver LTM pour chaque utilisateur lorsque le rappel semble incorrect. Une fonctionnalité de personnalisation qui ne peut pas être désactivée devient un cauchemar pour le support. Séparez également les « faits de profil » des « résumés de conversation » dans le stockage afin qu’une mauvaise rotation de résumés ne puisse pas effacer des préférences durables. Lors du passage de InMemoryStore à Postgres, répétez les opérations de sauvegarde et de restauration des tables de mémoire de la même manière que vous répétez les restaurations de points de contrôle — la perte de LTM correspond à une amnésie visible pour l’utilisateur, et non simplement à un problème technique.

Apprendre à l’extracteur ce que signifie « important »

Les prompts qui indiquent « enregistrer tout ce qui est utile » génèrent du contenu inutile. Préférez une liste de contrôle fermée : identité, préférences durables, projets en cours, contraintes strictes, ainsi que des rappels fournis explicitement par l’utilisateur. Tout le reste reste dans le point de contrôle du fil de discussion. Montrez à l’extracteur trois paires d’échanges bruités par rapport au JSON de mémoire minimal souhaité, et relancez ces tests dans l’environnement CI lorsque les prompts changent. La qualité de la mémoire à long terme est le produit d’un prompt et d’un schéma ; traitez-la en conséquence.

Les opérateurs devraient également s’entraîner à corriger les « mémoires erronées » : supprimer par clé, extraire à nouveau à partir d’une phrase corrigée, et s’assurer que le nœud de conversation ne cite plus ce fait obsolète. Sans cet entraînement, la mémoire à long terme devient un tiroir uniquement destiné à l’écriture de données inutiles.