AI agents get dumber at half their advertised memory
Builders call the halfway mark the dumb zone, where performance fades long before memory runs out.
Lightspeed Venture Partners · The $20,000 Robot, AI's Essay Wars & Loop Engineering Explained | LightworkAn artificial intelligence agent starts to get noticeably dumber when it has filled only about half of its advertised memory.
That is the warning from investor Namdi Eregbulum, known as Nomi, who works with companies building long-running AI helpers. The helpers work in loops, and on each lap they stack up more material: their own outputs, instructions from the user, and new data from the outside world.
A common model might advertise room for 270,000 tokens, the chunks of text and data it can consider at once, and some advertise one million. But builders have a nickname for what happens around the halfway mark: the dumb zone.
The model has not run out of room, yet its work gets worse. The result, Eregbulum says, is that the amount of memory builders can actually use is quite small.
Questions people ask
Why do agents fill up memory so fast?
A chatbot handles one question at a time, but an agent saves everything as it works. Each step adds its own draft, the user's instructions, plus tool results, web pages, and error messages. After dozens or hundreds of steps, that running transcript gets very long.
Does buying a model with a bigger memory fix the dumb zone?
It helps, but it does not eliminate the problem. A one-million-token window holds more history than a 270,000-token window, yet accuracy still tends to fall as the input grows. Builders still try to keep the important facts early and small, and clear out the rest.
How do builders keep an agent out of the dumb zone?
They summarize old steps, save key facts to files, and start fresh runs with only the summary. Many systems also split work across smaller subagents with short memories. The goal is to use a small, clean slice of memory rather than the whole advertised window.