A Temporal Knowledge Graph Architecture for AI Agent Memory.pdf

AI agents need reliable long-term memory for real-world tasks.
Existing memory models (e.g., RAG, vector stores) struggle with:
Temporal reasoning
Data evolution tracking
Precise retrieval from long histories
Zep addresses this using a temporal knowledge graph approach

1. 10 May 2025
Green University Of Bangladesh
Jahidul Islam - 221002504
Md. Al-Imran - 221002470
Dr. Faiz Al Faisal
Associate Professor and Director of GCITS
Green University Of Bangladesh
Paper Presentation on:
Zep: A Temporal Knowledge Graph Architecture for
AI Agent Memory
01
Presented By: Presented To:

2. Literature Review
Methodology
Result Analysis
Table of Contents
02
10 May 2025
Green University Of Bangladesh

3. ZEP : A TEMPORAL KNOWLEDGE GRAPH ARCHITECTURE FOR
AI AGENT MEMORY
Published on
Jan 2025
Citations 11+
Authors
Preston Rasmussen, Pavlo Paliychuk,
Travis Beauvais, Jack Ryan, Daniel
Chalef
Affiliations
ZEP AI
Publisher arXiv
Literature Review
03
10 May 2025
Green University Of Bangladesh

4. 04
Motivation & Problem Statement:
AI agents need reliable long-term memory for real-world tasks.
Existing memory models (e.g., RAG, vector stores) struggle with:
Temporal reasoning
Data evolution tracking
Precise retrieval from long histories
Zep addresses this using a temporal knowledge graph approach.
MOTIVATION & PROBLEM
10 May 2025
Green University Of Bangladesh

5. 05
Temporal Knowledge Graph:
Memory is structured as a bi-temporal graph with nodes
and relations.
Stores facts as triples: (subject, predicate, object)
Each fact is stored with:
Valid Time: When the fact is true in the world
Transaction Time: When it was recorded
Enables querying over time-aware, evolving knowledge.
CORE IDEA - TKG
Fig. 01: How Knowledge Graph works
10 May 2025
Green University Of Bangladesh

6. TEMPORAL KNOWLEDGE GRAPH IN ACTION
06
Fig. 02: TKG in Action.

7. 07
Three key components:
Input Processing: Extracts and links structured triples from agent conversations.
Graphiti Engine:
Manages the evolving TKG
Supports fast, incremental updates
Hybrid Retrieval: Combines symbolic querying with neural reasoning (e.g., LLMs)
ARCHITECTURE OVERVIEW
10 May 2025
Green University Of Bangladesh

8. 08
Built for:
Real-time updates
Bi-temporal indexing
Conflict resolution in memory
incremental ingestion (no full reprocessing)
Scalable to millions of events
Provides temporal indexing and automatic resolution of contradictory data
Time-aware retrieval across dynamic enterprise contexts.
GRAPHITI – THE MEMORY ENGINE
10 May 2025
Green University Of Bangladesh

9. Gaps & Limitations
09
Gaps in existing Works
Poor temporal reasoning (e.g., MemGPT,
RAG)
No structured, evolving memory
Slow or inefficient long-term recall
Zep Addresses These With
Bi-temporal knowledge graph
Fast, time-aware retrieval
Scalable, real-time updates
Limitations
Poor temporal reasoning (e.g., MemGPT,
RAG)
No structured, evolving memory
Slow or inefficient long-term recall
10 May 2025
Green University Of Bangladesh

10. Results & Contributions
Deep Memory Retrieval (DMR)
LongMemEval
Evaluation Metrics:
Accuracy
Latency
Token Usage
Datasets:
10
Tab. 1. Significant improvements of ZEP
Bi-temporal memory graph for accurate, time-
aware agent recall
Graphiti engine enables fast, scalable updates
Outperforms traditional vector/RAG memory in
precision and speed
Contributions:
Task Zep MemGPT LongMem Improvement
Deep Memory
Retrieval
94.80% 93.40% 92.70%
+1.4% vs
MemGPT
LongMemEval
Accuracy
84.50% 66.00% 65.50%
+18.5% vs
MemGPT
Latency
(ms/query)
28 ms 280 ms 310 ms ~90% faster
10 May 2025
Green University Of Bangladesh

11. 11
Use Cases:
Customer support, policy tracking, multi-agent collaboration
APPLICATIONS & FUTURE WORK
Future Directions:
Shared memory across agents
Human-in-the-loop memory editing
Integration into large enterprise systems
10 May 2025
Green University Of Bangladesh

12. Green University Of Bangladesh
12
Referecnces
[1] Zep AI, "Zep: A Temporal Knowledge Graph Architecture for Agent
Memory," arXiv preprint arXiv:2501.13956, Jan. 2025. [Online]. Available:
https://arxiv.org/abs/2501.13956
10 May 2025

13. THANK YOU
Green University Of Bangladesh
13