Agent Memory and Context Management: Conversation History, Summarization, and State Persistence

Agents that handle a single request are trivial. Real agents handle multi-turn conversations: user asks question 1, agent responds, user follows up with question 2 based on the previous answer. Memory systems maintain conversation history and enable agents to reference past interactions.

Conversation Buffer Memory

from langchain.memory import ConversationBufferMemory

memory = ConversationBufferMemory()

# First turn
memory.save_context(
    {"input": "What is 2+2?"},
    {"output": "2+2 equals 4"}
)

# Second turn - agent can see history
memory.load_memory_variables({})
# Output: {"history": "Human: What is 2+2?\nAI: 2+2 equals 4"}

Summarization Memory

For long conversations, buffer memory grows unbounded. Summarization compresses history.

from langchain.memory import ConversationSummaryMemory

memory = ConversationSummaryMemory(llm=llm)

# After many turns, conversation is summarized
summary = memory.buffer  # Compressed conversation summary

Vector Memory

Store important facts in a vector store for long-term recall.

from langchain.memory import VectorStoreBackedMemory

memory = VectorStoreBackedMemory(
    vectorstore=vector_store,
    k=5  # Retrieve top 5 relevant memories
)

Conclusion

Memory systems enable agents to handle multi-turn conversations and maintain long-term context. Buffer memory is simple but unbounded; summarization compresses; vector memory enables retrieval. Choosing the right memory strategy depends on conversation length and user expectations. Next: error handling and recovery in agents.