Vector Store Retriever Memory

1. What is Vector Store Retriever Memory?

VectorStoreRetrieverMemory stores conversation information as embeddings in a vector database and retrieves only the most relevant past information when needed.

It is semantic memory, not chronological memory.

---

2. Why does it exist?

Other memories fail when:

• Conversations are very long

• Old but important facts must be recalled

• Exact wording doesn’t matter, meaning does

Vector retriever memory solves:

• Long-term memory

• Semantic recall

• Scalable memory storage

In short:

Remember by meaning, not by order.

---

3. Real-world analogy

Think of:

• Buffer memory → chat log

• Summary memory → meeting notes

• Vector memory → search engine for your past conversations

You “search” your memory by meaning.

---

4. Minimal working example (Gemini + FAISS)

from langchain_google_genai import ChatGoogleGenerativeAI
from langchain.memory import VectorStoreRetrieverMemory
from langchain.chains import ConversationChain
from langchain.vectorstores import FAISS
from langchain.embeddings import GoogleGenerativeAIEmbeddings
import os

llm = ChatGoogleGenerativeAI(
    model="gemini-2.5-flash",
    api_key=os.getenv("GEMINI_API_KEY")
)

embeddings = GoogleGenerativeAIEmbeddings(
    model="models/embedding-001",
    google_api_key=os.getenv("GEMINI_API_KEY")
)

vectorstore = FAISS.from_texts([], embeddings)

retriever = vectorstore.as_retriever(
    search_kwargs={"k": 2}
)

memory = VectorStoreRetrieverMemory(
    retriever=retriever
)

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

conversation.invoke("My name is John and I live in Toronto")
conversation.invoke("I work at Google")
conversation.invoke("Where do I live?")

---

5. What gets stored?

Each interaction is stored as:

• Text

• Converted to embeddings

• Saved in vector DB

Example stored chunks:

"My name is John and I live in Toronto"
"I work at Google"

---

6. How does retrieval work?

When a new question comes in:

1. Question is embedded

2. Vector DB finds similar past messages

3. Top-k results are injected into prompt

Only relevant memory is used.

---

7. Key parameters

Parameter	Meaning
k	How many past memories to retrieve
Chunk size	Granularity of memory
Vector DB	FAISS, Chroma, Pinecone

---

8. Comparison with other memories

Memory Type	Recall Style
Buffer	Chronological
Summary	Compressed
Entity	Fact-based
KG	Relationship-based
Vector	Semantic

---

9. Common mistakes

❌ Assuming it remembers everything ❌ Not chunking memory properly ❌ Using it without relevance filtering

Vector memory retrieves similar, not exact matches.

---

10. When should you use it?

Use VectorStoreRetrieverMemory when:

• You need long-term memory

• Conversations are large

• Semantic recall matters

Avoid when:

• Short chats

• Precise step ordering matters

---

11. One-line mental model

VectorStoreRetrieverMemory = semantic search over past conversations