ConversationBufferWindowMemory

1. What is ConversationBufferWindowMemory?

ConversationBufferWindowMemory stores only the last K interactions of a conversation.

Instead of remembering everything, it remembers a sliding window of recent messages.

---

2. Why does it exist?

ConversationBufferMemory grows forever → high token cost.

ConversationBufferWindowMemory solves this by:

• Limiting memory size

• Keeping only recent context

• Preventing token explosion

In short:

Remember what matters most: the recent conversation.

---

3. Real-world analogy

Talking to a human who:

• Clearly remembers the last few minutes

• Forgets older details unless repeated

That’s exactly how window memory behaves.

---

4. Minimal working example (Gemini)

from langchain_google_genai import ChatGoogleGenerativeAI
from langchain.memory import ConversationBufferWindowMemory
from langchain.chains import ConversationChain
import os

llm = ChatGoogleGenerativeAI(
    model="gemini-1.5-flash",
    api_key=os.getenv("GEMINI_API_KEY")
)

memory = ConversationBufferWindowMemory(
    k=2  # remember last 2 exchanges
)

conversation = ConversationChain(
    llm=llm,
    memory=memory,
    verbose=True
)

conversation.invoke("My name is John")
conversation.invoke("I live in Toronto")
conversation.invoke("What is my name?")

---

5. What actually gets remembered?

With k=2, memory keeps:

Human: I live in Toronto
AI: ...
Human: What is my name?

❌ "My name is John" is forgotten because it fell outside the window.

So the model may fail to answer correctly.

---

6. Key parameter: k

k value	Meaning
k=1	Only last exchange
k=2	Last 2 user–AI turns
k=5	Last 5 turns
k=0	No memory

---

7. Inspecting the memory buffer

print(memory.buffer)

Shows only the most recent messages, not the full history.

---

8. Comparison with other memory types

Memory Type	Stores	Token Cost
Buffer	Full chat	High
Window	Last K turns	Medium
Summary	Summarized history	Low

---

9. Common mistakes

❌ Setting k too small and losing important facts ❌ Assuming it “understands importance” (it doesn’t) ❌ Using it for long-term facts (names, preferences)

---

10. When to use it

Use ConversationBufferWindowMemory when:

• You need short-term context

• Conversations are medium length

• Cost control matters

• You don’t need long-term recall

---

11. One-line mental model

ConversationBufferWindowMemory = last K messages only