Your AI Agent has an Amnesia Problem

The essay pulls three Chain of Thought conversations into one argument: enterprise AI agents usually fail because of how they are given information, and almost never because of the model. Each guest supplies one layer of a three-layer data stack for agents.

Michel Tricot, co-founder of Airbyte, covers context management in episode 52. In his live demo, a query through Airbyte’s context store retrieved 45 Gong calls in about a minute. Without it, the agent paged through every call, because Gong’s API cannot filter by user, and spent three minutes and 30,000 extra tokens. Richmond Alake, Director of AI Developer Experience at Oracle, covers memory architecture in episode 53. He maps agent memory onto human working, semantic, episodic and procedural memory. Sudhir Hasbe, President and CPO of Neo4j, covers knowledge graphs in episode 55.

Key points

• Context poisoning: an agent wired straight to APIs spends its capacity paginating and managing data access, and its reasoning degrades.
• Alake’s two principles are “think memory first” and “don’t delete, forget.” Decay by recency, relevance and importance keeps an audit trail that a hard delete would lose.
• Hasbe’s accuracy figures for complex enterprise queries: about 60 to 65% for pure vector RAG, above 90% with a graph layer, and 95 to 97% with reasoning tooling on top.
• The essay says 78% of enterprises have agent pilots, cites a survey in which only 14% have scaled one to production, and notes a Gartner prediction that over 40% of agentic AI projects will be canceled by 2027.

Disclosure: Episodes 52, 53 and 55 were sponsored by Galileo, the Season 3 presenting sponsor (episode 52 read at 1:30, episode 53 thanks at 0:28, episode 55 read at 1:37).