엔터티 및 요약 메모리 전략

A plain buffer stores every message verbatim. In long chats this quickly exceeds the model context window and wastes tokens on irrelevant chatter.

Two smarter strategies fix this: summary memory compresses the past, and entity memory tracks specific facts about people and things.

Summary memory uses an LLM to keep a running, condensed summary of the conversation. Instead of replaying 50 turns, the agent reads a few sentences capturing the gist.

• Keeps token usage roughly constant
• Preserves long-term context
• Loses fine-grained wording

LangChain's ConversationSummaryMemory calls the LLM after each turn to update the summary. You pass it the same model the agent uses.

As turns happen you save context; reading the variables returns the compressed history rather than raw turns.

ConversationSummaryBufferMemory keeps recent turns verbatim AND summarizes older ones once a token limit is crossed. Recent context stays sharp while old context is compressed.

Entity memory extracts named things — people, projects, places — and stores a fact sheet for each. When an entity reappears, the agent recalls exactly what it knows.

This is ideal for personal assistants that must remember user preferences.

The entity memory uses the LLM to detect entities and maintain a per-entity store.

Each entity accumulates facts. Asking about an entity loads only its relevant summary into the prompt.

Pick based on the use case:

• Buffer: short, exact conversations
• Summary: long chats where the gist matters
• Summary Buffer: long chats needing recent precision
• Entity: assistants tracking facts about specific subjects

Summary and entity memory make extra LLM calls on every turn to update their state. That adds cost and latency.

Use cheaper, faster models for the memory-update step than for the main agent reasoning when possible.

For sophisticated agents you can combine memories with CombinedMemory, e.g. a summary for flow plus entity memory for facts. Just ensure their output keys do not collide.

Test your understanding of advanced memory strategies.

You learned memory strategies beyond raw buffers:

• Summary memory compresses history with the LLM
• Summary buffer keeps recent turns exact and summarizes the rest
• Entity memory tracks facts about named subjects
• Each adds LLM calls — balance cost vs. recall
• Combine memories for richer agents

Choosing the right strategy keeps agents both knowledgeable and efficient.