Super Data Science: ML & AI Podcast with Jon Krohn 985: The Four Types of Memory Every AI Agent Needs, with Richmond Alake
111 snips
Apr 21, 2026 
Richmond Alake, Oracle’s Director of AI Developer Experience and creator of the 100 Days of Agent Memory, discusses agent memory systems. He explains four memory types, why RAG alone falls short, and how memory fits into the broader agent stack. He also covers engineering tradeoffs, memory-first harnesses, and Oracle’s unified AI database tools.
AI Snips
Chapters
Books
Transcript
Episode notes
Semantic Memory Holds Domain Facts And Institutional Knowledge
- Semantic memory is durable world or institutional knowledge used to ground agents in domain facts.
- Richmond points to enterprise institutional knowledge as semantic memory for agents to mirror employee know-how.
Working Memory Is The Model Context Window
- Working memory maps to the LLM context window and is the short-term info used in real time during an interaction.
- Richmond equates working memory directly with the context window the model uses for immediate reasoning.
Don't Treat RAG As Complete Memory
- Use RAG for retrieval but extend beyond it: update, consolidate, resolve conflicts, and forget memories actively.
- Richmond warns RAG doesn't handle memory updates or conflict resolution needed for production agents.
