AI agents lose what they learn when they give notes to a successor
Six models learned less in a chain of eight agents with 300 character notes. They matched one agent only after about four lives.
Researchers studied agents that give work to a successor as a short written summary. They used 56 interactive worlds with hidden rules. One agent played eight lives and kept its full history. A chain of eight agents each played one life and got a 300 character note. All six models learned less in the chain.
Where the loss comes from
The team divided the loss into three parts. A change of agent caused none. A cut to 300 characters caused a smaller loss. Repeated cuts caused a larger loss.
What helped
Each note can come from the full record of earlier lives. The same 300 characters then recovered 70% to 82% of the loss. The loss was the same in all model dimensions. Three instructions did not help:
- Record proof.
- Change the note that the writer got.
- Keep a structured ledger.
What is not known
The abstract does not name the six models. It says that a rule that takes some lives to see is not found by the lives that got no note.
Insight: The fix for lost lessons is the store that leaks secrets
An earlier test (17 October 2026) found that agent memory keeps signs of a secret after the message is removed. This paper finds that notes decrease lessons, and that the full record gets them back. A team must select between a full record that holds secrets and a short note that loses lessons. This looks like a choice to make before launch.
Sources
Posted