Collections

Best Chain of Thought Episodes on Agent Memory

How agents remember — memory stores, context engineering, and what breaks without them.

Every team that ships an agent hits the same wall around month two. It works inside one session and falls apart across many. What decides that is the memory layer: what the agent keeps, and what it’s allowed to forget. Most teams build it last, after the failure has already cost them something. These are the ones I point people to when they get there.

  1. 1 EP 53 59 min Transcript Agent Memory: The Last Battleground in the AI Stack | Richmond Alake, Oracle Richmond Alake, Oracle Richmond Alake makes the case that memory engineering is its own discipline, separate from prompt and context engineering, and lands on the rule that separates production systems from prototypes: don’t delete, forget.
  2. 2 EP 52 44 min Transcript Context Poisoning is Killing Your AI Agents: How to Stop it Michel Tricot, Airbyte Airbyte’s Michel Tricot on context poisoning — agents failing because the data reaching them is wrong, not because the model is. He demos a query that burned 30,000 extra tokens before a context store got in front of it.
  3. 3 EP 61 53 min Transcript The AI Framework Era Is Over: Why Context Is the Moat | Jerry Liu Jerry Liu, LlamaIndex Jerry Liu argues context quality is the moat that compounds while the scaffolding around it gets absorbed into the model. The last stretch is his read on agent memory as persistent context.
  4. 4 EP 33 40 min Transcript Mastering Multi-Agent Systems | MongoDB’s Mikiko Chandrasekhar Mikiko Chandrasekhar, MongoDB MongoDB’s pitch to be the memory store for agents, and what the data layer has to take on once an agent’s state outlives the conversation.
  5. 5 EP 42 53 min Transcript Architecting AI Agents: The Shift from Models to Systems | Aishwarya Srinivasan Aishwarya Srinivasan, Fireworks AI Aishwarya Srinivasan on the move from prompt engineering to context engineering, and why a production agent needs memory orchestrated alongside its models and tool calls rather than bolted on.

More collections

All collections →

New here? Start with five → Browse by topic →