LangChain Model Memory Not Working: Why Gemini Forgets Names and How to Fix It

If your LangChain agent with MemorySaver fails to recall your name when using Gemini models, the problem is not the memory system. It is a model behavior difference. The memory is working correctly. The model interprets certain prompts as knowledge recall tasks rather than conversational lookups. To get consistent recall, rephrase your question to "Do you know my name?" or "Do you remember my name?" instead of "What's my name?" or switch to a model like Claude 3.5 Sonnet or Llama 3.2 that handles this prompt more naturally.

The Full Answer

This issue appears when you follow the official LangChain agent tutorial but use Google Gemini models instead of Anthropic Claude. The tutorial at https://python.langchain.com/docs/tutorials/agents/ uses anthropic:claude-3-5-sonnet-latest. That model behaves differently from Gemini when answering certain memory-related prompts.

How MemorySaver Works

MemorySaver from langgraph.checkpoint.memory stores the full message history for a given thread ID. When you call agent_executor.stream() with the same thread_id, LangGraph automatically retrieves the previous messages and includes them in the context window for the next request. You do not need to manually fetch history and prepend it.

Here is the standard setup from the official tutorial (adapted for Gemini):

import os
from langchain_tavily import TavilySearch
from langchain.chat_models import init_chat_model
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.messages import HumanMessage
from langgraph.prebuilt import create_react_agent

os.environ.get('TAVILY_API_KEY')
search = TavilySearch(max_result=2)
tools = [search]

os.environ['HTTP_PROXY'] = 'http://127.0.0.1:7890'
os.environ['HTTPS_PROXY'] = 'http://127.0.0.1:7890'
os.environ.get('GOOGLE_API_KEY')
model = init_chat_model('gemini-2.5-flash', model_provider='google-genai')

memory = MemorySaver()
agent_executor = create_react_agent(model, tools, checkpointer=memory)
config = {'configurable': {'thread_id': 'agent003'}}

# First call
for step in agent_executor.stream(
    {'messages': [HumanMessage('Hi! I am Bob!')]}, config, stream_mode='values'
):
    step['messages'][-1].pretty_print()

# Second call
for step in agent_executor.stream(
    {'messages': [HumanMessage("What's my name?")]}, config, stream_mode='values'
):
    step['messages'][-1].pretty_print()

With Gemini 2.5 Flash, the second call returns something like: "I'm sorry, I don't have memory of past conversations. Could you please tell me your name again?"

Why "What's my name?" Fails

As confirmed by community member Ajeet Verma on Stack Overflow, the failure is not a LangGraph memory bug. It is a model behavior issue specific to Gemini. When you ask "What's my name?", Gemini interprets this literally as a knowledge recall task. Since the model does not have an internal persistent memory store that persists across API calls, it defaults to "I don't know your name."

When you ask "Do you know my name?" or "Do you remember my name?", the model interprets this more conversationally. It looks at the immediate chat history included in the same request and correctly extracts "Bob."

Here are the three runs with Gemini 2.5 Flash using the exact same code but different prompts:

1st run: prompt "What's my name?"

================================ Human Message =================================

Hi! I am Bob!
================================== Ai Message ==================================

Hello Bob! How can I help you today?
================================ Human Message =================================

What's my name?
================================== Ai Message ==================================

I'm sorry, I don't have memory of past conversations. Could you please tell me your name again?

2nd run: prompt "Do you know my name?"

================================ Human Message =================================

Hi! I am Bob!
================================== Ai Message ==================================

Hello Bob! How can I help you today?
================================ Human Message =================================

Do you know my name?
================================== Ai Message ==================================

Yes, your name is Bob.

3rd run: prompt "Do you remember my name?"

================================ Human Message =================================

Hi! I am Bob!
================================== Ai Message ==================================

Hello Bob! How can I help you today?
================================ Human Message =================================

Do you remember my name?
================================== Ai Message ==================================

Yes, I do, Bob!

Solution 1: Rephrase the Question

If you must use Gemini, change your application logic to ask memory questions in a conversational style. Instead of "What's my name?", use "Do you know my name?" or "Do you remember my name?" This works because the model treats these as context lookups rather than knowledge retrieval.

Solution 2: Switch to a Different Model

If you cannot control the user prompts, switch to a model that handles direct recall questions more reliably. The official tutorial uses anthropic:claude-3-5-sonnet-latest. Community testing shows that llama3.2:latest from Ollama also works well.

Here is the same code adapted for Ollama with Llama 3.2:

import os
from langchain_tavily import TavilySearch
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.messages import HumanMessage
from langgraph.prebuilt import create_react_agent
from langchain_ollama import ChatOllama
from dotenv import load_dotenv

load_dotenv()
os.environ.get('TAVILY_API_KEY')
search = TavilySearch(max_result=2)
tools = [search]

model = ChatOllama(
    model="llama3.2:latest", temperature=0
)

memory = MemorySaver()
agent_executor = create_react_agent(model, tools, checkpointer=memory)

# Same thread_id for continuity
config = {"configurable": {"thread_id": "agent003"}}

# First turn
for step in agent_executor.stream(
    {"messages": [HumanMessage("Hi! I am Bob!")]}, config, stream_mode="values"
):
    step["messages"][-1].pretty_print()

# Second turn
for step in agent_executor.stream(
    {"messages": [HumanMessage("what's my name?")]}, config, stream_mode="values"
):
    step["messages"][-1].pretty_print()

Output with Llama 3.2:

================================ Human Message =================================
Hi! I am Bob!
================================== Ai Message ==================================
Tool Calls:
....
================================= Tool Message =================================
Name: tavily_search
....
================================== Ai Message ==================================
Your name is Bob! I've found multiple individuals with the name Bob, including Bob Marley, B.o.B, and Bob Iger. Is there a specific Bob you're interested in learning more about?

Notice that Llama 3.2 correctly answers "Your name is Bob!" to the direct question "what's my name?" without any prompt rephrasing.

Manual History Injection (Not Recommended)

One user attempted to manually fetch history and prepend it to the new message:

state = memory.get(config)
history = state["channel_values"]["messages"]
new_input = HumanMessage("What's my name?")
full_context = {"messages": history + [new_input]}

for step in agent_executor.stream(full_context, config, stream_mode='values'):
    step['messages'][-1].pretty_print()

This approach did not change the outcome with Gemini. The model still forgot the name. This confirms that the issue is not about whether history is present in the context window. It is about how Gemini interprets the prompt.