Player not loading? Watch on YouTube
This walkthrough introduces Memory Bases for Langflow 1.10, which the speaker describes as a way to retain chat context across conversations. An embedding model converts conversation history into vectors for storage in a vector database. The speaker distinguishes this from a knowledge base built around files.
The setup uses a paralegal AI agent with OpenAI models. The presenter creates a memory base, chooses an embedding model, and sets the ingestion batch size to one so each exchange triggers a job. Optional LLM preprocessing uses custom instructions to decide which information to retain. In this example, the prompt extracts and summarizes legal context and includes a scoring rule that accepts or rejects messages.
A breach-of-contract example tests whether the agent can recall the parties and their roles. The presenter then asks about travel to France and checks that the memory base contains no travel details.
Session filtering controls whether retrieval uses the current chat session or information across sessions. With filtering disabled, a greeting brings up the earlier legal case. A later chat with filtering enabled answers that Acme Corp has not appeared in its context.
