From Data Models to Mind Models: Designing AI Memory at Scale

The podcast explores agentic memory architectures, focusing on how AI agents can retain state and context across interactions. Vasilije Markovich, founder of Cognee, discusses the evolution of memory solutions from simple vector stores to complex graph-vector layers that support business rules and ontologies. He details the importance of both permanent and session memory, highlighting use cases where low latency is critical. The conversation covers practical aspects of memory implementation, including storage media, metadata attributes for filtering, and strategies for managing the decay of memory utility over time. Markovich also addresses the challenges of memory isolation in multi-agent systems and the role of human feedback in improving agent performance.

Outlines

Part 1: Introduction, Background

Part 2: Memory Architectures, Use Cases

Part 3: Technical Implementation, Storage

Part 4: Management, Governance

Part 5: Optimization, Practical Applications

Part 6: Lessons, Future Outlook

Sign in to continue reading, translating and more.

Continue

Data Engineering Podcast

Part 1: Introduction, Background

Introduction to Agentic Memory Architectures with Vasilije Markovich

Vasilije's Journey: From Business to Cognitive Science and Dynamic Schemas

Part 2: Memory Architectures, Use Cases

Agentic Memory: Permanent vs. Session Memory Architectures

Use Cases for Different Memory Types: Data Silos, Agentic Tasks, and Edge Cases

Part 3: Technical Implementation, Storage

Storage Media for Memories: From Markdown Files to Vector Databases

Structuring Memory Information: Metadata, Embeddings, and Agent-Specific Graphs

Addressing Memory Decay: From Recency to Reinforcement Learning Techniques

Triggering Memory Storage and Retrieval: Tool Calls and Pseudo-SQL

Part 4: Management, Governance

Memory Sharing and Isolation: Multi-Tenancy and Public Memories

Schema Evolution: Managing Changes in Memory Structure and Content

LLM Responsibility vs. Curated Data: Nature vs. Nurture in AI

Part 5: Optimization, Practical Applications

Self-Improvement and Fine-Tuning: The Role of Memory and Human Feedback

Adding Memory to Systems: From Excel to Data Warehouses and Beyond

Unexpected Applications of Agentic Memory: From Theology to Cybersecurity

Part 6: Lessons, Future Outlook

Lessons Learned: Modularity, Communication, and Ignoring the Noise

When Not to Invest in Memory: Prompts, Structured Outputs, and Postgres

Future Plans: Decision Traces, Memory Transformations, and Time Representation

The Future of Memory: Complex Embeddings and Policy Logic Engines

From Data Models to Mind Models: Designing AI Memory at Scale

Data Engineering Podcast

Part 1: Introduction, Background

00:11Introduction to Agentic Memory Architectures with Vasilije Markovich

Introduction to Agentic Memory Architectures with Vasilije Markovich

01:14Vasilije's Journey: From Business to Cognitive Science and Dynamic Schemas

Vasilije's Journey: From Business to Cognitive Science and Dynamic Schemas

Part 2: Memory Architectures, Use Cases

04:42Agentic Memory: Permanent vs. Session Memory Architectures

Agentic Memory: Permanent vs. Session Memory Architectures

09:43Use Cases for Different Memory Types: Data Silos, Agentic Tasks, and Edge Cases

Use Cases for Different Memory Types: Data Silos, Agentic Tasks, and Edge Cases

Part 3: Technical Implementation, Storage

12:47Storage Media for Memories: From Markdown Files to Vector Databases

Storage Media for Memories: From Markdown Files to Vector Databases

16:11Structuring Memory Information: Metadata, Embeddings, and Agent-Specific Graphs

Structuring Memory Information: Metadata, Embeddings, and Agent-Specific Graphs

23:44Addressing Memory Decay: From Recency to Reinforcement Learning Techniques

Addressing Memory Decay: From Recency to Reinforcement Learning Techniques

26:06Triggering Memory Storage and Retrieval: Tool Calls and Pseudo-SQL

Triggering Memory Storage and Retrieval: Tool Calls and Pseudo-SQL

Part 4: Management, Governance

28:46Memory Sharing and Isolation: Multi-Tenancy and Public Memories

Memory Sharing and Isolation: Multi-Tenancy and Public Memories

32:04Schema Evolution: Managing Changes in Memory Structure and Content

Schema Evolution: Managing Changes in Memory Structure and Content

35:45LLM Responsibility vs. Curated Data: Nature vs. Nurture in AI

LLM Responsibility vs. Curated Data: Nature vs. Nurture in AI

Part 5: Optimization, Practical Applications

37:58Self-Improvement and Fine-Tuning: The Role of Memory and Human Feedback

Self-Improvement and Fine-Tuning: The Role of Memory and Human Feedback

40:15Adding Memory to Systems: From Excel to Data Warehouses and Beyond

Adding Memory to Systems: From Excel to Data Warehouses and Beyond

43:47Unexpected Applications of Agentic Memory: From Theology to Cybersecurity

Unexpected Applications of Agentic Memory: From Theology to Cybersecurity

Part 6: Lessons, Future Outlook

47:08Lessons Learned: Modularity, Communication, and Ignoring the Noise

Lessons Learned: Modularity, Communication, and Ignoring the Noise

50:36When Not to Invest in Memory: Prompts, Structured Outputs, and Postgres

When Not to Invest in Memory: Prompts, Structured Outputs, and Postgres

52:05Future Plans: Decision Traces, Memory Transformations, and Time Representation

Future Plans: Decision Traces, Memory Transformations, and Time Representation

54:30The Future of Memory: Complex Embeddings and Policy Logic Engines

The Future of Memory: Complex Embeddings and Policy Logic Engines