buffer-window-chat-memory

BufferWindowChatMemory — Sliding Window Conversation Memory

BufferWindowChatMemory is a variant of BufferChatMemory that only retains the last k messages in a conversation. This is useful when you want to keep the most recent context short and focused, instead of passing the entire conversation back to the model.

---

Concept

• Stores messages in a BaseChatHistory.
• When reading memory (load_variables or chat_buffer_window), it returns only the most recent k messages.
• Perfect for keeping prompts lightweight while still giving the model near-term context.

---

API

• k: integer (default 5). How many messages to retain.
• save_context(message): add a new message to memory.
• chat_buffer_window(): returns a list of the last k messages.
• load_variables(): expose the windowed buffer as a string, ready to inject into the next prompt.
• clear(): reset all stored messages.

---

Example

from tinygent.core.datamodels.messages import TinyHumanMessage, TinyChatMessage
from tinygent.memory import BufferWindowChatMemory

memory = BufferWindowChatMemory()
memory.k = 3  # keep only the last 3 messages

memory.save_context(TinyHumanMessage(content="Hello"))
memory.save_context(TinyChatMessage(content="Hi there!"))
memory.save_context(TinyHumanMessage(content="What's up?"))
memory.save_context(TinyChatMessage(content="All good!"))
memory.save_context(TinyHumanMessage(content="Thanks!"))

print("Windowed memory:", memory.chat_buffer_window())
print("As variables:", memory.load_variables())

Output:

Windowed memory: [TinyChatMessage("All good!"), TinyHumanMessage("Thanks!")]
As variables: {'last_3_messages': '...'}

---

When to Use

• Keep prompts short and within token limits.
• Provide only recent context to the model.
• Ideal for long conversations where full history is unnecessary.