How can I add 'Memory' to a LangChain agent so it remembers user preferences across sessions?

All answers to this question.

To achieve long-term persistence, you need to move away from in-memory buffers and use something like RedisChatMessageHistory or a SQL-backed history. LangChain has built-in integrations for Redis, MongoDB, and Postgres. You’ll need to store the session_id in your database. When the user returns, you initialize the chain with that specific session_id to fetch the previous chat history. For specific preferences like "vegetarian," it’s actually better to use a "Profile" agent that summarizes the conversation and saves "entities" into a Vector Store like Pinecone.

   Answered 2025-01-25 by Sandra Mitchell

Should we be worried about the context window getting too bloated if we pull in a month's worth of chat history every time the user says hi?

   Answered 2025-01-26 by George Lopez

• George, absolutely. You should use ConversationSummaryBufferMemory. It keeps the last few messages in raw format for immediate context but summarizes the older parts of the conversation. This way, the agent remembers the "vegetarian" preference from three weeks ago via the summary, but the prompt stays lean and cost-effective. You basically get the best of both worlds.

     Commented 2025-01-27 by William Wright

I suggest looking into Zep. It’s a long-term memory layer designed specifically for LLM apps that handles the summarization and embedding of past chats automatically for you.

   Answered 2025-01-28 by Susan Foster

• I second the Zep recommendation! It integrates perfectly with LangChain and handles the heavy lifting of memory management so you can focus on the agent's actual logic.

     Commented 2025-01-29 by Matthew Robinson