My users are complaining that the bot "forgets" context after about five or six exchanges. I am using the Buffer Memory node in Flowise, but it doesn't seem to hold up. Should I be using a different memory type like Zep or a specific DynamoDB integration for long-term retention? I need the bot to remember user preferences across different login sessions.
3 answers
Buffer Memory is only meant for short-term "sliding window" context. If you want true long-term persistence in Flowise, you absolutely should integrate Zep or use a persistent database. I personally use the Postgres Chat Memory node. This allows the bot to query past interactions even if the user logs out and comes back a week later. You can set a "Session ID" that maps to your user's unique ID in your app. This ensures that the LLM always has the relevant historical context without blowing through your token limit by sending the entire history every time.
Are you worried about the token costs increasing as the conversation history grows longer over several weeks?
I found that the DynamoDB node is the most reliable for AWS-based deployments if you want to scale horizontally.
Great tip, Pamela. Using an external managed service for memory makes the Flowise instance itself stateless, which is a best practice for modern cloud architecture and deployment.
Definitely, Justin. That’s why I’m looking at the "Conversation Summary Memory" node in Flowise. It uses a smaller LLM to summarize the history, which keeps the token count low while still preserving the "gist" of the previous sessions.