Heavybit Podcasts Ep. #20, Exploring AI Memory with Vasilije Markovic of Cognee
Aug 21, 2025
In this fascinating discussion, Vasilije Markovic, a data engineering and cognitive science expert from Cognee, delves into the world of AI memory. He tackles challenges like 'context rot' and emphasizes the need for effective data organization to enhance AI’s decision-making capabilities. Markovic shares insights on AI's practical applications across industries like construction and education, while humorously illustrating the complexities of context and relational data. The conversation also touches on the intricate relationship between creativity and technology in today's landscape.
AI Snips
Chapters
Books
Transcript
Episode notes
Why RAG Alone Often Fails
- RAG (vector search) often returns noisy, overly broad chunks rather than the precise fact you need.
- AI memory organizes entities and relationships so models get only the context that matters.
Multi-Layer Memory Mirrors Human Language
- Human memory stores language across layers (words, phrases, sentences) with cross-connections.
- Cognee models use multi-layer memory to extract entities and link context like an ontology.
Ground LLMs With Ontologies
- Combine LLM-extracted entities with ground-truth RDF ontologies to avoid hallucinated facts.
- Use explicit ontologies to mark reliable facts versus LLM-generated context.
