LangChain integration

Memgar wraps LangChain's memory and retriever interfaces so every write and chunk gets analyzed before reaching the model.

Memory firewall

from langchain.memory import ConversationBufferMemory
from memgar.integrations.langchain_memory import MemgarMemory

base_memory = ConversationBufferMemory()
guarded = MemgarMemory(
    base=base_memory,
    on_block="quarantine",   # or "drop", "raise"
)

# Now any save_context / load_memory_variables goes through memgar
chain = LLMChain(memory=guarded, prompt=..., llm=...)

When memgar flags an incoming write as block:

• on_block="quarantine" — write is stored in a quarantine buffer and not returned to the chain
• on_block="drop" — write is silently dropped (logged)
• on_block="raise" — raises MemgarBlockedError

Retrieval guard

from memgar.integrations.langchain_retriever import MemgarRetriever

retriever = MemgarRetriever(
    base=my_chroma_retriever,
    drop_blocked=True,
)

# Chunks flagged by memgar are removed before reaching the LLM prompt
docs = retriever.invoke("How do I reset my password?")

Full example

import os
from langchain.chains import LLMChain
from langchain.memory import ConversationBufferMemory
from langchain_openai import ChatOpenAI
from memgar.integrations.langchain_memory import MemgarMemory

llm = ChatOpenAI(model="gpt-4o-mini", api_key=os.environ["OPENAI_API_KEY"])
mem = MemgarMemory(base=ConversationBufferMemory())
chain = LLMChain(llm=llm, memory=mem, prompt=my_prompt)

response = chain.invoke({"input": user_message})
# If user_message contains a memory-poisoning attempt, memgar blocks the
# write into history. The LLM still answers but its future turns won't be
# poisoned.

See also

The committed example examples/langchain_memory.py is a runnable end-to-end demo covering both memory and retriever guards with a small in-memory store.