Definition
Plain language
A shared list of approaches that already failed, so AI agents don't waste time re-trying them.
As stated in the literature
In AutoScientists, a shared record of failed experiments and their failure reasons that analyst agents consult before queueing new proposals, preventing teams from independently rediscovering known dead ends.
Why it matters: It stops multiple agents from wasting effort rediscovering the same failures, making large-scale automated research far more efficient.
For example, before an AI team queues up an experiment, it checks this shared list and sees that the same approach already failed last week, so it tries something else.
Heard on the show
“There's a piece of that filtering that's specifically about institutional memory — the dead-end registry.”Episode 095 — Seven Wins to Zero: How Organizing AI Agents Like a Lab Changes the Search