langchain/docs/modules/memory/key_concepts.md

Key Concepts

Memory

By default, Chains and Agents are stateless, meaning that they treat each incoming query independently. In some applications (chatbots being a GREAT example) it is highly important to remember previous interactions, both at a short term but also at a long term level. The concept of "Memory" exists to do exactly that.

Conversational Memory

One of the simpler forms of memory occurs in chatbots, where they remember previous conversations. There are a few different ways to accomplish this:

• Buffer: This is just passing in the past N interactions in as context. N can be chosen based on a fixed number, the length of the interactions, or other!
• Summary: This involves summarizing previous conversations and passing that summary in, instead of the raw dialouge itself. Compared to Buffer, this compresses information: meaning it is more lossy, but also less likely to run into context length limits.
• Combination: A combination of the above two approaches, where you compute a summary but also pass in some previous interactions directly!

Entity Memory

A more complex form of memory is remembering information about specific entities in the conversation. This is a more direct and organized way of remembering information over time. Putting it a more structured form also has the benefit of allowing easy inspection of what is known about specific entities. For a guide on how to use this type of memory, see this notebook.