ConversationKGMemory

1. What is ConversationKGMemory?

ConversationKGMemory builds a knowledge graph (KG) from a conversation.

Instead of storing plain facts, it stores relationships in the form:

(subject) — (relation) — (object)

Example:

(John) — lives_in — (Toronto)
(John) — works_at — (Google)

---

2. Why does it exist?

Entity memory stores flat summaries:

“John lives in Toronto and works at Google.”

That becomes hard to reason over.

KG memory solves:

• Multi-hop reasoning

• Relationship queries

• Structured long-term memory

In short:

Remember facts as relationships, not sentences.

---

3. Real-world analogy

Think of:

• Entity memory → notebook paragraphs

• KG memory → a graph database (like Neo4j)

KG memory allows asking:

• “Where does John live?”

• “Who works at Google?”

• “Which city does John live in?”

---

4. Minimal working example (Gemini)

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

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

memory = ConversationKGMemory(
    llm=llm
)

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 and where do I work?")

---

5. What does it store internally?

You can inspect the graph:

print(memory.kg.get_triples())

Example output:

[
 ('John', 'lives_in', 'Toronto'),
 ('John', 'works_at', 'Google')
]

This is a true graph structure, not free text.

---

6. How does it work internally?

Each conversation turn:

1. LLM extracts entities

2. LLM infers relationships

3. Triples are added to the graph

4. Relevant triples are injected into the next prompt

So the LLM:

• Builds the graph

• Queries the graph

• Reasons over relationships

---

7. Key characteristics

Feature	KG Memory
Stores	Subject-predicate-object
Structure	Graph
Multi-hop reasoning	✅
Token usage	Low
Conversation flow	❌

---

8. Entity Memory vs KG Memory

Feature	Entity	KG
Data type	Text facts	Triples
Reasoning	Weak	Strong
Structure	Flat	Graph
Query flexibility	Medium	High

---

9. Common mistakes

❌ Expecting it to remember chat wording ❌ Using it for short conversations ❌ Using it without structured questions

KG memory shines when relationships matter.

---

10. When should you use it?

Use ConversationKGMemory when:

• Relationships are important

• You need long-term structured memory

• You want reasoning across facts

Avoid when:

• Chat nuance matters

• You need verbatim recall

• Conversations are short

---

11. One-line mental model

ConversationKGMemory = conversation → knowledge graph