Shared AI agent memory can poison itself with stale lessons, Airbnb's CTO warns

Ahmad Al-Dahle, CTO of Airbnb, argued on The TWIML AI Podcast that a shared memory for AI agents can go bad as it grows. Old or wrong lessons get saved, and agents then find and use them. Airbnb is building such a system, called Project Everest. He said he does not know whether it will hold up at scale.

What was said

Everest collects know-how from employees as they work with AI. Agents can then use things that never made it into the documentation. Al-Dahle described it as effectively a company brain.

He named the main risk as context poisoning. His example was a plain-text notes file. One person uses the system, the team extracts a lesson and saves it to the file. Al-Dahle said "the model's like very good at finding the thing that's old or wrong."

He also warned against giving a model too much information at once. His preferred approach is to give just enough to take one step, then let the model ask for more.

Why it matters

Our reading: collecting knowledge is the easy part, and keeping it correct is the hard part. A memory that only grows, with no way to retire old notes, risks handing agents bad advice, and every agent that reads it inherits the error.

For teams building or buying agent memory, the questions are practical. How are old notes found and removed? What test shows the memory helps, rather than just adding volume? Al-Dahle's pricing task is one model of such a test.

The other side

Al-Dahle was open about the limits. He said Everest has not solved everything and that it is early. He said he is not convinced the team has the right architecture yet, and he doubts that vector search, a common way to look up stored text, will scale company-wide.

He also left the central worry open. Asked whether Everest might be damaged as more people use it, he said the team does not know. He raised the possibility that the system would need a rewrite.