Memory Types in LangChain: Buffer, Summary, and Vector — When to Use Each

More memory isn't always better memory — sometimes it's just more billed tokens without improving the quality of the response.

Buffer memory: everything, exactly as it happened

Conversation Buffer Memory stores the full conversation history exactly as it occurred, keeping every message sequentially — though it can become inefficient in long conversations due to token limits and cost considerations. Conversation Buffer Window Memory retains a set number (2*k) of conversation pieces as history, by fixing a value of k — a variant that bounds growth without losing the most recent context.

Summary memory: a summary instead of a full transcript

Conversation Summary Memory condenses the past historical conversation log into summaries for storage, where the memory used is the summary, not the full dialogue data. Conversation Summary Buffer Memory keeps conversation history data without exceeding a `max_token_limit`, and for the parts that exceed that limit, the information gets extracted and summarized — a hybrid between buffer and summary that balances recent fidelity with compression of older content.

Vector memory: retrieving only what's relevant

Vector Store-Backed Memory uses vector similarity to retrieve semantically relevant documents from a store, improving context retrieval. For specific factual retrieval, VectorStoreRetrieverMemory is used — instead of loading the entire history or a generic summary, this variant actively searches for which part of the history is relevant to the current question.

The practical 2026 recommendation

Start with ConversationBufferMemory or ConversationBufferWindowMemory for basic use cases, and if memory becomes a bottleneck or too costly, explore ConversationSummaryMemory or ConversationBufferWindowMemory with a smaller k value — the recommended progression goes from simple to complex, not the other way around, avoiding over-engineering before confirming it's actually needed.

How to choose for your use case

For short customer support conversations: buffer window memory with a small k. For long-term assistants with extensive conversations: summary buffer memory. For agents that need to remember specific facts mentioned many turns ago, without loading the entire history: vector memory. The right decision depends on whether what matters is the complete recent conversation, summarized general context, or specific isolated facts.